WORKSPACE_MEMORY.md

# Workspace Memory

## ATEC2026 L0 Task E

- Scope:
  - Current priority is L0 Task E tabletop manipulation in `/home/ubuntu/Documents/01Proj/13atec/ATEC2026_Simulation_Challenge`; do not switch to L1 unless the user asks.
  - 2026-06-07 官网文件上传式提交记录：XSA-ACT 上传后线上得分 `6.00`，构建号 `BUILD2001385_2038204482147323904_20260607142442_bae481`，任务“桌面整理 / Piper”，用时约 `00:11:22.62`。`get_action_spec not found` warning 不是扣分原因，已在 `demo/solution_act.py` 和 `submissions/task_e_xsa_act_20260607_upload/solution.py` 增加 `get_action_spec(self): return None` 消除 warning；更新后的 `solution.py` SHA256 `a08790e7512c914e5b78132303e1c6098c8aba3b01a373933e6ddc617fa3a485`。
  - 2026-06-07 PCA/GraspGen-style submit 分支复测：`ATEC_PCA_OBJECTS=3,2,1`、seed11、`demo/solution_pca.py` 跑满 5200 steps 得 `0.00`，日志 `logs/pca_submit_eval/pca_submit_seed11_20260607.log`，视频 `logs/videos/task_e_pca_submit_eval/pca_submit_seed11_20260607.mp4`。失败现象：RGB-D/PCA 中心估计可生成，但 submit-side IK/finger 实际位置长期追不上 desired finger center，object3/object2/object1 都未入篮。当前不要把 `solution_pca.py` 当作打榜提交。
  - 2026-06-07/08 为线上试榜准备两个 ACT seed2 checkpoint 包，均已瘦身为推理 checkpoint（约 64MB）并通过 `AlgSolution()` 初始化：`submissions/task_e_act_seed2_45000_20260607_upload`，zip SHA256 `4b5d1178db3060883db4400f24eaa04b11b259a5179a220a7a622d7347ea2895`，`policy_act.pt` SHA256 `d9badbf17b0f29000c4fb21c3a2a89f66f7450e3ec1cb7362e733f4a4a7a0cb2`；历史本地 seed11 得 `18`、seed12 得 `0`，线上用户回报得 `12.00`。`submissions/task_e_act_seed2_50000_20260607_upload`，zip SHA256 `1d777cfa95ee40d92afafd70f5e3cbfd3ee6477f0b036ed9d459c94237858118`，`policy_act.pt` SHA256 `8ac243e8427c54704a0445fa8f0321450618fa6c2d2d7296db027c9c5fe1e2c9`；历史本地 seed11 得 `15`、seed12 得 `6`，线上用户回报得 `6.00`。两个包都用更新后的 `solution.py` SHA `a08790e...` 和无 `IPython` 的 `act/detr/transformer.py` SHA `fa290d...`。结论：seed2 checkpoint 方差大，不能视为稳定 18 分方案。
  - 2026-06-08 准备旧 ACT seed1 best “抽奖候选”提交包：`submissions/task_e_act_seed1_best_20260608_upload`，zip `submissions/task_e_act_seed1_best_20260608_upload.zip`，zip SHA256 `e265ebe77b989a968cf4607ea70d4f3adc2322902538828b37fe84877f5b1391`，`policy_act.pt` SHA256 `282614de9673dc01e229557448e380b6a02c196b5c3953cec77b2aebd2305a8e`，`solution.py` SHA256 `a08790e7512c914e5b78132303e1c6098c8aba3b01a373933e6ddc617fa3a485`。来源是旧 seed1 best（原备份 SHA `76bbef2c81c85fc6cc60448476ff09075ec8bc97217c2d2fb70a3253fd3e307d`），已瘦身只保留 `norm_stats`、`model_args`、`ema_agent` 并通过 `AlgSolution()` 初始化。历史本地独立 run1 得 `18`、run2/run3 得 `6`，线上尚未提交；若用户还能多次试榜，这是当前已打包候选里最值得下一次提交的包。
  - 2026-07-19 赛后开源归档：模型仓库为 `https://huggingface.co/Datawhale/atec2026-task-e-act-seed1-best`，已上传 `policy_act.pt` 与完整提交 zip；数据/日志仓库为 `https://huggingface.co/datasets/Datawhale/atec2026-task-e-reproducibility`。复现训练数据使用 filtered HDF5：`datasets/atec_task_e_obj321_servo_100demos/trajectory_filtered.hdf5`，大小约 33GiB，SHA256 `4b85980e72b8e76261ad037bfba97b4795210e3f77cdb5b71479846b9230ac02`；HF 上按 100MiB 拆为 `data/final_100demos_filtered_split_100m/trajectory_filtered.hdf5.part-0000..0336`，配套 `RESTORE_FILTERED_DATASET.sh`、`SHA256_FILTERED_ORIGINAL.txt`、`SHA256SUMS_FILTERED_SPLIT.txt`。上传采用暂存目录 `/data/Data14TB/01Proj/ATEC2026/hf_task_e_filtered_100m` 和 user service `atec-task-e-hf-filtered-100m-upload.service`，入口 `/data/Data14TB/01Proj/ATEC2026/upload_task_e_filtered_100m_large_folder.py`，设置 `HF_HUB_DISABLE_XET=1` 与 `upload_large_folder(num_workers=8)` 断点续传。2026-07-19 已完成远端校验：337/337 分片齐全、`README.md`、`RESTORE_FILTERED_DATASET.sh`、两个 SHA256 清单和 `logs/task_e_logs_20260520_20260609.tar.zst` 均存在；上传暂存目录可清理，但原始 `trajectory_filtered.hdf5` 仍保留，除非用户另行确认。
  - 2026-06-08 用户已提交 `task_e_act_seed1_best_20260608_upload/policy_act.pt` 到官网，线上出分 `15.00`；当天出分/提交次数紧张，用户表示此后只剩一次提交机会。该结果强于 XSA 线上 `6`、seed2_45000 线上 `12`、seed2_50000 线上 `6`，目前是线上最佳。最后一次提交必须有本地明确优于 15 的证据，否则不要用掉。
  - 最后一投风险判断，2026-06-08：历史日志显示 XSA final 和 seed2_50000 通常 object_2/object_3 入篮，object_1 失败；seed2_45000 本地 seed11 object_1/2/3 全入篮但 seed12 全崩，线上 `12` 说明隐藏评估仍丢至少一个物体。PCA/GraspGen-style submit 分支当前 seed11 为 `0`，完整 GraspNet/AnyGrasp 不符合短期单文件稳定提交条件。不要在最后一次机会中使用未通过 submit-style 本地评估的 PCA/GraspNet 方案。
  - 最后一投 hybrid 试验，2026-06-08：构建了 `submissions/task_e_act_seed1_hybrid_obj1_pca_20260608_upload`，思路是 ACT seed1/XSA 先做前两物体，之后切 `solution_pca.py` 的 object_1-only 几何补救。初始化修复包括提交目录内复制 `solution_pca.py`、Piper URDF `third_party/Agilex-College/piper/handpose_det/models/modified_piper_without_camera.urdf`，并把 `solution_pca.py` URDF 搜索改成优先当前提交目录。结果失败：seed1-best hybrid seed11 `score=6.00`，只有 object_3 inside，视频 `logs/videos/task_e_hybrid_eval/hybrid_obj1_pca_seed11_20260608.mp4`；XSA final hybrid 固定 ACT step 1320 切 PCA 后 seed11 `score=3.00`，object_1/2/3 均不 inside，视频 `logs/videos/task_e_hybrid_eval/hybrid_xsa_fixed1320_obj1_pca_seed11_20260608.mp4`。目录内已加 `DO_NOT_SUBMIT.md`；不要用该 hybrid 包做最后一次线上提交。
  - 2026-06-08 下午继续冲分负结果：新增实验文件 `demo/solution_hybrid_obj1_pca.py`（仅 `current_score>=15` 后切 object_1 PCA）和 `demo/solution_dual_act.py`（两套 ACT 同步观测、score 阈值后切副策略），都还不是提交方案。`logs/eval_task_e_hybrid_xsa_score15_obj1_pca_seed11_20260608.log` 得 `0.00`；`logs/eval_task_e_hybrid_seed1best_score15_obj1_pca_seed12_20260608.log` 得 `0.00`；`logs/eval_task_e_hybrid_seed1_95000_score15_obj1_pca_seed11_20260608.log` 得 `0.00`，视频 `logs/videos/task_e_hybrid_eval/hybrid_seed1_95000_score15_obj1_pca_seed11_20260608.mp4`。双 ACT：`logs/eval_task_e_dualact_seed1best_seed2_45000_switch12_seed11_20260608.log` 得 `0.00`；`logs/eval_task_e_dualact_seed1_95000_seed2_45000_switch12_seed11_20260608.log` 在 `score=15`、primary_steps=900 切副策略，但最终仍 `15.00`，object_1 外、object_2/3 内。结论：score=15 后的 PCA/双 ACT 接力没有把 object_1 补入篮，不能消耗最后一次提交机会。
  - 2026-06-08 继续 hybrid 冲 18 的新增边界：`runs/.../seed1/checkpoints/95000.pt` trace seed11 显示 step 300 `score=6`、step 600 `score=12`，object_2/object_3 已入篮而 object_1 仍在桌面，日志 `logs/trace_task_e_seed1_95000_seed11_20260608.log`。据此测试 `score=12 + min_act_policy_steps=600` 切 object_1 PCA：`logs/eval_task_e_hybrid_seed1_95000_score12_step600_obj1_pca_seed11_20260608.log` 仍 `15.00`，object_1 最终 `(0.988,0.336,0.872)` 外，object_2/3 内；打开旧 object_1 rescue 的 `logs/eval_task_e_hybrid_seed1_95000_score12_step600_obj1_pca_rescue_seed11_20260608.log` 退到 `12.00` 且未触发 hybrid switch。已给 `demo/solution_hybrid_obj1_pca.py` 加显式 `ATEC_HYBRID_FORCE_ENV_STEPS` 计数和切换原因打印，避免只依赖 ACT 内部 `_ts`/score 回调；已给 `demo/solution_pca.py` 增加 `ATEC_PCA_OBJ1_DIRECT_RESCUE=1` 实验入口，并修正 `_build_object1_drag_rescue()` 的明显 bug：旧版 `start_y=TABLE_CENTER_Y+0.085` 会在 object_1 位于 `y≈0.26..0.34` 时从方块前方错过，现改为从当前方块中心正 y 侧动态生成 lane，且 lane x 也围绕当前检测中心。由于 2026-06-08 16:15-16:45 机器 GPU/Isaac 并发很重，多条 force-env eval 在 Isaac 启动后迟迟进不到有效 step，已手动停止；这些新 direct-rescue 补丁尚未形成 >15 本地证据，不得提交。
  - 2026-06-08 seed1 checkpoint sweep：对 `runs/act-task-e-rgb-obj321-servo-100demos-seed1/checkpoints` 粗筛 seed11。结果：`35000=0`、`45000=12`、`50000=0`、`55000=6`、`65000=0`、`70000=0`、`75000=6`、`80000=6`、`90000=9`、`95000=15`、`best_loss=0`，对应日志 `logs/eval_task_e_seed1_*_20260608_sweep*.log`。扩测 `95000.pt`：seed12 `6`、seed13 `6`，因此 95000 不是比线上 seed1-best 15 更稳的替换候选。当前最佳仍是线上已有 `task_e_act_seed1_best_20260608_upload` 的 `15.00`，不要提交 95000 或 dual/hybrid。
  - 2026-06-04 已整理中文交接文档：`docs/ATEC2026_TaskE_规则物理配置与提交说明_20260604.md`；官方 README 快照下载到 `docs/sources/ATEC2026_Simulation_Challenge_official_readme_20260604.md`。文档结论：当前是官方 `ATEC-TaskE-Piper` / Isaac Lab v2.3.2 环境，Piper solver、物体质量、摩擦/contact 参数来自本地官方源码；香蕉滑落更像二指夹爪接触保持和 submit-side primitive 问题，不能靠修改 judge 物理提交，物理增强只适合诊断。
  - 2026-06-04 fresh clone 官方仓库到 `/tmp/ATEC2026_Simulation_Challenge_official_check`，HEAD `dbe7c251f680b02f357a6db67430b18d3ba45ea1`。核对结果：`piper.py`、`object.py`、Task E `terrain.py`、`rewards.py`、`terminations.py` 与本地 SHA 完全一致；`task_e/env_cfg.py` 只有 action group 字段名差异，物体/篮筐/相机/Y-band/physics material 逻辑一致。官方最新版新增 `demo/server.py` 的 `/get_action_spec` 和 custom action 示例，这属于动作接口可选自定义，不是允许改 judge 物理。
  - 2026-06-04 再次 `git fetch origin` 官方仓库，`origin/main` 仍停在 `dbe7c25`，没有新 commit。GitHub issues/PR 网页核对：未发现公开 issue 提到 Task E/Piper/grasp/physics/friction/banana 滑落；`Piper` 搜索命中的是 PR #2 eye-in-hand camera for B2Piper/B2WPiper/G1，不是 Task E Piper；PR #5 是 participant action spec，PR #8 是 Tron2A joint configuration。已同步唯一低风险兼容修复：给本地 `demo/server.py` 添加官方 `/get_action_spec` endpoint；若 solution 无该方法则返回 `{}`，保持默认 action config。
  - Official Task E robot is `ATEC-TaskE-Piper`; the newer `Tron2AWheel` / `Tron2ALegged` models are not relevant for this tabletop task.
- 调试复盘总览，2026-05-30：
  - 当前结论：可部署基线仍是 XSA-ACT 的 `demo/policy_act.pt`，SHA256 `c3eacae3f1b0ec8ccde9fdc05d680fff119cc2ca70e1f79e4ddd5610c4bbfa93`。它不是满分稳定方案，但本地证据仍强于 seed2 ACT、GraspNet/SAM3 直接控制和 pi0.5/OpenPI 分支。不要用 pi0.5 当前 checkpoint 覆盖 `demo/policy_act.pt`。
  - 机器人/任务身份：Task E 使用 `ATEC-TaskE-Piper`，本地 env 注册在 `source/atec_rl_lab/atec_rl_lab/tasks/task_e/__init__.py`，Piper 配置来自 `TaskEEnvPiperCfg` / `PIPER_CFG`。PiPER 是 AgileX/松灵机械臂系列；本任务是桌面级固定 Piper，不是 L1 的 Tron/Unitree 轮腿/足式机器人。
  - 数据生成主线：官方/初始 scripted oracle 对 object_3 相对可用，但 object_1/object_2 失败，根因不是相机，而是 USD/root 几何中心、Piper gripper TCP、夹爪实际接触几何和运输/投放 primitive 不匹配。最终通过 per-object calibrated primitive 和闭环 servo 生成了三物体 `3 2 1` 成功数据。核心数据目录是 `datasets/atec_task_e_obj321_servo_100demos`，过滤后 HDF5 是 `datasets/atec_task_e_obj321_servo_100demos/trajectory_filtered.hdf5`。
  - ACT seed1 主线：用上述 100 demos 训练普通 ACT 后，第一次评估失败是 `demo/solution_act.py` home 阶段理解错 action：Task E action 是 `(joint_target - default_joint_pos) / 0.5`，不是速度/PD delta。修复 home action 后普通 ACT 可部分成功，但 `best_loss.pt`/`final.pt` 不稳定，训练 loss 与任务成功率弱相关。旧 seed1 best 已备份为 `demo/policy_act.seed1_best_backup_20260522_012719.pt`。
  - XSA-ACT 主线：用户提供的 `xsa_update_20260520/lerobot_policy_act_xsa.zip` 是 LeRobot `act_xsa` 插件，原始例子不是 ATEC 8D Piper 直接可跑。已在 `source/atec_rl_lab/atec_rl_lab/train/act/act/detr/transformer.py` 和 `demo/act/detr/transformer.py` 增加 `--use_xsa` 适配；`scripts/act/train_task_e.py` 保存 `model_args`，`demo/solution_act.py` 能按 checkpoint 自动加载 `use_xsa`。XSA final 直接独立评估 seeds `11/12/13` 曾得 `15/15/15`，日志 `logs/eval_task_e_xsa_final_seed11_recheck.log`、`logs/eval_task_e_xsa_final_seed12_recheck.log`、`logs/eval_task_e_xsa_final_seed13_recheck.log`；但部署后复查有 seed11 `9`、seed12 `0` 的接触/非确定性波动。仍然是当前最强可部署基线。
  - ACT seed2/resume 主线：seed2 从头训练并加了 resume 支持，相关改动在 `scripts/act/train_task_e.py` 和 `scripts/act/run_task_e_pipeline.sh`，支持 `--resume_checkpoint` / `--resume_iter`、`RESUME_CHECKPOINT` / `RESUME_ITER`，并保存 optimizer/scheduler/EMA。seed2 过程中有局部 snapshot 看似高分，例如 `45000.pt` seed11 `18`，但 seed12 `0`；最终 `final.pt` seed11 `0`，`best_loss.pt` seed11 `9`。结论：seed2 不部署，且不要把单 seed 高分当作可靠证据。
  - GraspNet/TunTunClaw 主线：用户给的 OpenClaw/Datawhale TunTunClaw 参考已接到 `third_party/tuntunclaw/`，GraspNet checkpoint 是 `third_party/tuntunclaw/temp/logs/log_rs/checkpoint-rs.tar`；桥接脚本是 `scripts/graspnet_task_e/tuntun_adapter.py` 和 `scripts/graspnet_task_e/run_graspnet_pick.py`。依赖在 `/home/ubuntu/envs/genmanip-isaac5-py311/bin/python` 下打通，包括 `open3d==0.19.0`、本地 `graspnetAPI`、`transforms3d==0.4.2`、Blackwell `TORCH_CUDA_ARCH_LIST=12.0` 的 `pointnet2` CUDA extension。GraspNet 能给候选，但原始候选不能直接解决真实 Isaac 接触；object_3 可通过 SAM3/GraspNet+真实 finger/staged transport 成功，object_1/object_2 仍在运输阶段滑落/偏出，不能作为最终控制器。
  - SAM3 主线：SAM3 checkpoint 位于 `/home/ubuntu/Documents/01Proj/sam3d_gs/submodule/Prompt-Inpaint/checkpoints/sam3.pt`，独立 Python 在 `/home/ubuntu/Documents/01Proj/sam3d_gs/.venv/bin/python`，不要装进 Isaac env。新增 `scripts/graspnet_task_e/sam3_segment_image.py` 做 RGB prompt 分割；`run_graspnet_pick.py` 支持 `--mask_provider sam3`、`--sam3_prompt`、`--basket_center_release`。重要结论：SAM3 分割不是主要瓶颈，object_1/2/3 的 mask 都能做对；早期失败是多候选 mask 选错，后来用 world-band/RGB-D 点数筛选修复。真正瓶颈是 Piper 夹爪接触和运输阶段防滑。
  - GraspNet/SAM3 放置规则：用户提出“目标中心点上方、稳定后再松爪”是正确方向。已加入 transport-end logging、短稳定 hold、`--release_z`、`--basket_xy_tol`、`--basket_hold_steps`、`--basket_stable_steps`、`--basket_recovery_steps`、`--staged_transport`、`--transport_servo_fraction` 等参数。经验：不要长时间高空 hold，object 会 slip/bounce；object_3 用短稳定 hold 更好。object_1/object_2 仍不稳，不能只调松爪时机，需要重做 grasp/contact primitive。
  - pi0.5/OpenPI 主线：OpenPI repo 是 `/home/ubuntu/src/openpi-ebench-clean`，env 是 `/home/ubuntu/envs/openpi-pi05`，通用 checkpoint 是 `/home/ubuntu/projects/robotics_shared/checkpoints/openpi/pi05-ebench-generalist/pi05_generalist/params`。必须设置 `HF_LEROBOT_HOME=/home/ubuntu/projects/robotics_shared/datasets/lerobot`、`HF_HOME=/home/ubuntu/projects/robotics_shared/hf_cache`、`PYTHONPATH=/home/ubuntu/src/openpi-ebench-clean/src:/home/ubuntu/src/openpi-ebench-clean/packages/openpi-client/src`。不要在 OpenPI repo 里 `uv run`，会创建巨大 `.venv` 并重复下载。
  - pi0.5 数据转换：`scripts/pi05/convert_task_e_hdf5_to_lerobot.py` 把 ATEC HDF5 的 8D qpos/action 映射成 OpenPI fixed-base 16D padded 格式：`state/action.joints[:6]` 对应 Piper 6 arm joints，`state/action.gripper[:2]` 对应两个夹爪关节，其余补零；三路相机 `overlook/left/right` 暂时都写同一张 224x224 RGB。正式数据集是 `/home/ubuntu/projects/robotics_shared/datasets/lerobot/atec/task_e_obj321_servo_100demos_s5_224`，约 4.7G，stride=5，fps=10。
  - pi0.5 关键 action 约束：ATEC HDF5 里的 `actions` 已经是 env action `(joint_target - default_joint_pos) / 0.5`，不是绝对 qpos。因此所有 ATEC OpenPI config 都必须 `extra_delta_transform=False`。若使用 OpenPI 默认 delta transform，会再减一次 state，训练目标会错。
  - pi0.5 训练结果：20-demo 快速分支 `pi05_atec_task_e_20demos_s5_224` 2k step 在 seeds `11/12/13` 全 `0`，checkpoint 已删，只保留 `logs/pi05_eval_20260522074137` 和视频 `logs/videos/task_e_pi05_eval/pi05_pi05_atec_task_e_20demos_s5_224_seed11_20260522074137.mp4`。100-demo 正式分支 `pi05_atec_task_e_s5_224` 训练 29999 step，checkpoint `.../pi05_atec_task_e_s5_224/atec_task_e_pi05_s5_224_30k_20260522065716/29999` 约 42G，评估 `logs/pi05_eval_20260522112701`：seed11 `0`，seed12 `0`，seed13 `6`，mean `2.0`，视频 `logs/videos/task_e_pi05_eval/pi05_pi05_atec_task_e_s5_224_seed11_20260522112701.mp4`。结论：pi0.5 当前分支不部署。
  - pi0.5 失败原因判断：这不是“pi0.5 模型弱”的直接证据，而是适配层很可能不对。当前桥接把 ATEC Piper 8D action 硬塞到 OpenPI EBench fixed-base 16D，三路相机全用同一张图，prompt 粗粒度，数据只有 scripted oracle 风格，且评估 max_steps=1500/action_repeat=5 与 ACT 1800-step 轨迹不完全等价。另一个强疑点是 OpenPI loader 日志里 `discrete_state_input=True`，而公开 Piper pi0.5 posttrain 经验提到 `discrete_state_input=False` 可避免 z-score norm 下 state token saturation。后续若继续 pi0.5，应优先做原生 Piper/OpenPI config、检查 q01/q99/norm stats、试 `discrete_state_input=False`，并考虑先试 pi0 而非 pi0.5。
  - pi0.5 旧分支核查追加，2026-06-08：已复查 `scripts/pi05/convert_task_e_hdf5_to_lerobot.py`、`demo/solution_pi05.py`、OpenPI `config.py` 和 `logs/pi05_train_task_e_s5_224_20260522065716.log`。旧 converter 将 8D Piper pad 成 EBench fixed-base `12 joints + 4 gripper`，其中真实 arm 在 `0:6`、真实 gripper 在 `12:14`，其余 8 个有效 EBench 维度全 0；随后 `PadStatesAndActions(model_action_dim=32)` 再 pad 到 32D。训练日志确认 `TokenizePrompt(... discrete_state_input=True)`，且三路图像 `head/hand_left/hand_right` 都是同一张 224 RGB。HDF5 原始数据无 NaN/Inf，动作确实是 env action，不是绝对 qpos；`extra_delta_transform=False` 是正确的。评估 `logs/pi05_eval_20260522112701` 显示 seed11/12 三物体基本停在桌面原 y 带，seed13 只 object_1 入篮，说明主要失败像是 embodiment/action-time bridge 和输入分布问题，而不是单纯“夹持滑落”。重启建议：保持 pi0.5 generalist 的 `action_dim=32` 以兼容预训练权重，但新增 ATEC-native transform，把真实 8D state/action 放在前 8 维并只输出前 8 维；设置 `Pi0Config(pi05=True, action_horizon=50, discrete_state_input=False)`；使用单相机真实 mask 或真实多相机，不要假三路；先用 stride=1 或 action_repeat=1/2，保存 1k/2k/5k/10k/15k 多 checkpoint 做 fixed-seed eval，再决定是否长训。
  - pi0.5 native8 重启分支，2026-06-08：新增 `scripts/pi05/convert_task_e_hdf5_to_lerobot_native8.py`、`demo/solution_pi05_native8.py`、OpenPI transform `/home/ubuntu/src/openpi-ebench-clean/src/openpi/policies/atec_piper_policy.py`，以及 configs `pi05_atec_task_e_native8_20demos_s1_224` / `pi05_atec_task_e_native8_100demos_s1_224`。设计意图：输入真实 8D qpos、输出真实 8D env action，模型内部仍 pad 到 32D 兼容 pi0.5 权重；只用 `video.base_camera_view` 一路真实图像，left/right wrist 用零图且 `image_mask=False`；`discrete_state_input=False`；stride=1/action_repeat=1，避免旧 stride=5 把抓取时序压坏。脚本路径：`scripts/pi05/run_pi05_native8_20demos_smoke.sh`、`scripts/pi05/run_pi05_native8_100demos_prepare.sh`、`scripts/pi05/run_pi05_native8_eval_checkpoint.sh`。由于根盘 `/` 只剩约 68G，native8 的 LeRobot 数据和 OpenPI assets/checkpoints 必须写到 `/data/Data4TB/01Proj/13atec/...`；不要再把新 42G checkpoint 写到 `/home/ubuntu/projects/...`。
  - pi0.5 native8 当前运行，2026-06-08 22:30 CST：旧 `atec-pi05-native8-20demo-smoke-20260608222155.service` 因同步图片写入太慢已主动停止；`atec-pi05-native8-20demo-smoke-20260608222519.service` 也主动停止，原因是 stride=1 转换约 1.5-2 小时，不适合先验 adapter 体检。converter 默认已改成 `image_writer_processes=5`、`image_writer_threads=10`。当前启动的是快速门控 user unit `atec-pi05-native8-20demo-s2-smoke-20260608223030.service`，执行 20-demo native8 stride=2 smoke（转换 -> norm stats -> 3k 训练 -> 自动评估 seeds 11/12/13），日志前缀 `logs/pi05_native8_convert_20demos_s2_224_*` / `logs/pi05_native8_norm_20demos_s2_224_*` / `logs/pi05_native8_train_20demos_s2_224_*` / `logs/pi05_native8_eval_20demos_s2_224_*`，数据目录 `/data/Data4TB/01Proj/13atec/robotics_shared/datasets/lerobot/atec/task_e_obj321_servo_20demos_native8_s2_224`。该分支是 adapter 体检，不是提交候选；评估使用 `CONFIG=pi05_atec_task_e_native8_20demos_s2_224 ACTION_REPEAT=2 scripts/pi05/run_pi05_native8_eval_checkpoint.sh <ckpt>`。必须在 seeds `11/12/13` 本地评估明显超过线上已得 `15.00`，才允许消耗最后一次官网提交机会。
  - pi0.5 native8 快速 sanity，2026-06-08 22:39 CST：20-demo s2 转换仍约 50 分钟，因此先主动停掉 `atec-pi05-native8-20demo-s2-smoke-20260608223030.service`，新增 `pi05_atec_task_e_native8_4demos_s2_224` / `scripts/pi05/run_pi05_native8_4demos_s2_sanity.sh`，并启动 `atec-pi05-native8-4demo-s2-sanity-20260608223911.service`。该分支只用 4 条 demo、stride=2、1k steps，训练后自动 `ACTION_REPEAT=2` 评估 seeds 11/12/13；目的只是尽快确认 native8 action bridge / OpenPI websocket / Isaac rollout 是否能让机械臂朝目标运动，不代表最终分数。若 sanity 视频完全不动或乱动，先修 bridge/norm；若动作方向正确，再恢复 20-demo/100-demo。
  - pi0.5 native8 sanity 结果，2026-06-08 23:15 CST：4-demo s2 训练完成，loss 从 `0.5766` 降到约 `0.0405`，但评估 seeds `11/12/13` 全 `0.00`，视频 `logs/videos/task_e_pi05_native8_eval/pi05_native8_pi05_atec_task_e_native8_4demos_s2_224_seed11_20260608231046.mp4`，montage `logs/videos/task_e_pi05_native8_eval/pi05_native8_4demo_s2_seed11_montage_20260608231046.jpg`。视频显示机械臂基本停在 home 附近；离线 action probe 用 checkpoint `999` 对训练帧推理，输出不是全零（多维动作幅度约 1-3），说明 websocket/unnormalize 基本通，但 4 条数据太少/轨迹未泛化，不能作为提交依据。该 4-demo checkpoint 占 `83G` 且无价值，已删除；保留日志、数据和视频。
  - pi0.5 native8 当前门控运行，2026-06-08 23:19 CST：已把 native8 configs 的 `save_interval/keep_period` 改成只保存 final，避免 40GB 中间 checkpoint 拖慢训练。重新启动 20-demo s2 user unit `atec-pi05-native8-20demo-s2-smoke-20260608231948.service`，脚本 `scripts/pi05/run_pi05_native8_20demos_s2_smoke.sh` 会转换 20 demos stride=2 -> norm stats -> 3k train -> 自动 `ACTION_REPEAT=2` 评估 seeds 11/12/13。若仍是 `0/0/0` 或视频仍停在 home，下一步不要加数据，先修 action timing/home gate/bridge；若有明确移动和非零分，再跑 100-demo native8。
  - pi0.5 native8 20-demo s2 结果，2026-06-09 00:42 CST：`pi05_atec_task_e_native8_20demos_s2_224` 训练 3k 完成，loss 最低约 `0.0188`，checkpoint `/data/Data4TB/01Proj/13atec/openpi_atec_runs/checkpoints/pi05_atec_task_e_native8_20demos_s2_224/atec_task_e_pi05_native8_20demos_s2_224_3k_20260608231948/2999`。评估日志 `logs/pi05_native8_eval_20260609003633` / 汇总 `logs/pi05_native8_eval_20demos_s2_224_20260608231948.log`，seeds `11/12/13` 得分 `6/0/0`，mean `2.00`；seed11 只有 object_3 入篮，视频 `logs/videos/task_e_pi05_native8_eval/pi05_native8_pi05_atec_task_e_native8_20demos_s2_224_seed11_20260609003633.mp4`。结论：native8 bridge 已能产生部分有效策略，但 20 条数据不足以超过线上 `15.00`，不得提交。
  - pi0.5 native8 100-demo s2 门控，2026-06-09 00:46 CST：新增 OpenPI config `pi05_atec_task_e_native8_100demos_s2_224`（100 demos、stride=2/fps=25、`discrete_state_input=False`、10k step、final-only checkpoint、assets/checkpoints 在 `/data/Data4TB/01Proj/13atec/openpi_atec_runs`）和脚本 `scripts/pi05/run_pi05_native8_100demos_s2_gate.sh`。已启动 user unit `atec-pi05-native8-100demo-s2-gate-20260609004546.service`，流程为转换 `trajectory_filtered.hdf5` -> norm stats -> 10k train -> `ACTION_REPEAT=2 PORT=8023` 评估 seeds `11/12/13`。日志前缀：`logs/pi05_native8_convert_100demos_s2_224_20260609004546.log`、`logs/pi05_native8_norm_100demos_s2_224_20260609004546.log`、`logs/pi05_native8_train_100demos_s2_224_20260609004546.log`、`logs/pi05_native8_eval_100demos_s2_224_20260609004546.log`。只有本地多 seed 明显超过线上 `15.00` 才考虑最后一次官网提交。
  - pi0.5 native8 100-demo s2 结果，2026-06-09 05:49 CST：转换完成 `100` episodes / `84877` frames，数据目录 `/data/Data4TB/01Proj/13atec/robotics_shared/datasets/lerobot/atec/task_e_obj321_servo_100demos_native8_s2_224` 约 `3.9G`；norm stats 写到 `/data/Data4TB/01Proj/13atec/openpi_atec_runs/assets/pi05_atec_task_e_native8_100demos_s2_224/atec/task_e_obj321_servo_100demos_native8_s2_224`。10k 训练完成，checkpoint `/data/Data4TB/01Proj/13atec/openpi_atec_runs/checkpoints/pi05_atec_task_e_native8_100demos_s2_224/atec_task_e_pi05_native8_100demos_s2_224_10k_20260609004546/9999` 约 `42G`；loss 从 `0.4456` 降到末段约 `0.009-0.013`，但任务评估很差。评估日志 `logs/pi05_native8_eval_20260609054339` / 汇总 `logs/pi05_native8_eval_100demos_s2_224_20260609004546.log`，seeds `11/12/13` 得分 `3/0/0`，mean `1.00`，三个物体均未入篮；seed11 视频 `logs/videos/task_e_pi05_native8_eval/pi05_native8_pi05_atec_task_e_native8_100demos_s2_224_seed11_20260609054339.mp4`。结论：该 100-demo native8-s2 全量分支不得提交，也不能消耗最后一次官网机会；线上 `15.00` ACT seed1 best 仍是受保护最佳。
  - pi0.5 native8 LoRA 入口/启动，2026-06-09 09:51 CST：参考第三方 Pantheon Piper pi0.5 LoRA（119 条 Piper single-arm episodes、`discrete_state_input=False`、7500 steps、peak LR `5e-5`、LoRA 保先验）的经验，已把 OpenPI config `pi05_atec_task_e_native8_100demos_s2_224_lora` 修成真正 LoRA 门控：`paligemma_variant="gemma_2b_lora"`、`action_expert_variant="gemma_300m_lora"`、显式 `freeze_filter=Pi0Config(...).get_freeze_filter()`、`num_train_steps=7500`、`peak_lr=5e-5`、`ema_decay=None`、`discrete_state_input=False`。当前仍使用本地已有 `/home/ubuntu/projects/robotics_shared/checkpoints/openpi/pi05-ebench-generalist/pi05_generalist/params` 作为 weight loader，因为本机未发现本地 `pi05_base` 目录；这与 Pantheon 的 “pi05_base no warm-start” 不完全相同，若该分支失败，下一步可补下载/验证 `gs://openpi-assets/checkpoints/pi05_base/params` 分支。新增脚本 `scripts/pi05/run_pi05_native8_100demos_s2_lora_gate.sh`，复用已转换数据 `/data/Data4TB/01Proj/13atec/robotics_shared/datasets/lerobot/atec/task_e_obj321_servo_100demos_native8_s2_224` 和已有 norm stats，不重复转换；已启动 user unit `atec-pi05-native8-100demo-s2-lora-gate-20260609095114.service`，训练日志 `logs/pi05_native8_lora_train_100demos_s2_224_20260609095114.log`，service 外层日志 `logs/atec-pi05-native8-100demo-s2-lora-gate-20260609095114.service.log`，checkpoint 根目录 `/data/Data4TB/01Proj/13atec/openpi_atec_runs/checkpoints/pi05_atec_task_e_native8_100demos_s2_224_lora`，训练后自动 `ACTION_REPEAT=2 PORT=8024` 评估 seeds `11/12/13`。门槛不变：只有本地多 seed 明显超过线上已有 `15.00`，才考虑用最后一次官网提交机会；否则继续保护 ACT seed1 best。
  - pi0.5 native8 LoRA 结果，2026-06-09 10:16 CST：`atec-pi05-native8-100demo-s2-lora-gate-20260609095114.service` 已完成。训练 7500 steps 正常结束，loss 从 `0.4452` 降到末段约 `0.013-0.019`，final checkpoint `/data/Data4TB/01Proj/13atec/openpi_atec_runs/checkpoints/pi05_atec_task_e_native8_100demos_s2_224_lora/atec_task_e_pi05_native8_100demos_s2_224_lora_7500_20260609095114/7499` 约 `8.6G`。自动评估 `ACTION_REPEAT=2`，日志 `logs/pi05_native8_lora_eval_100demos_s2_224_20260609095114.log` / `logs/pi05_native8_eval_20260609101354`，seed11/12/13 全部 `0.00`，三个物体均未入篮；seed11 视频 `logs/videos/task_e_pi05_native8_eval/pi05_native8_pi05_atec_task_e_native8_100demos_s2_224_lora_seed11_20260609101354.mp4`。结论：参考 Pantheon 的 LR/步数/`discrete_state_input=False`/LoRA freeze 后仍失败，说明当前 OpenPI native8 bridge/数据分布/动作时序问题未被 LoRA 解决；该 LoRA 不得提交，也不应消耗最后一次官网机会。临时心跳 `atec-pi0-5-native8-lora-gate-monitor` 已删除。
  - pi0.5 LoRA 失败排查，2026-06-09 10:30 CST：网页核对 OpenPI issue #672 和 Pantheon 卡片后，当前 LoRA 配置中的 `freeze_filter`、`ema_decay=None`、`discrete_state_input=False` 是对的；`num_steps=10/25/50` 离线采样对训练帧误差没有明显改善，说明不是单纯 ODE 采样步数太少。新增 `demo/solution_pi05_native8.py` 环境变量 `ATEC_PI05_CHUNK_EXEC_STEPS`（默认仍 50，不影响旧行为），测试 `ATEC_PI05_CHUNK_EXEC_STEPS=5` seed11 仍 `0.00`，视频 `logs/videos/task_e_pi05_native8_eval/pi05_native8_lora_chunk5_seed11.mp4`，说明也不是简单 50-step 开环太长。更强根因：pi0.5 当前训练数据来自 `trajectory_filtered.hdf5`，该文件由 `scripts/act/filter_demos.py` 删除静止/hold 步，100 demos 从 `749560` 帧压到 `169700` 帧，约删掉 `77%`；未过滤数据 qpos diff p50 `1.5e-5`，过滤后 qpos diff p50 `0.00936`、max `0.419`，这会破坏 OpenPI/pi0.5 50-step action chunk 的真实时间语义。另一个疑点：state 是绝对 qpos，但 action 是 ATEC env action `(target-default)/0.5`，与公开 Piper LeRobot 数据常见的 joint-position action 语义不一致；后续 pi0.5 应优先改成未过滤数据 + 绝对 joint target action，再由输出 transform 转回 env action。
  - pi0.5 rawabs/action-expert 排查，2026-06-09 12:30 CST：已按上条建议实现 raw unfiltered + absolute target action 体检。关键改动：`scripts/pi05/convert_task_e_hdf5_to_lerobot_native8.py` 支持 `--action_mode absolute_target` 和 `--max_frames_per_episode`；`/home/ubuntu/src/openpi-ebench-clean/src/openpi/policies/atec_piper_policy.py` 支持 `AtecPiperOutputs(action_mode="absolute_target")`，推理时把模型 absolute joint target 转回 ATEC env action；`demo/solution_pi05_native8.py` 默认 `ATEC_PI05_CHUNK_EXEC_STEPS=10`，并新增可选 `ATEC_PI05_ZERO_NOISE=1`/`ATEC_PI05_ACTION_HORIZON`。1-demo 数据集为 `/data/Data4TB/01Proj/13atec/robotics_shared/datasets/lerobot/atec/task_e_obj321_servo_rawabs_1demo_native8_s10_224`（raw `trajectory.hdf5`，1 episode，stride=10，250 frames，absolute target action）。
  - pi0.5 rawabs 结果，2026-06-09：新增 configs `pi05_atec_task_e_native8_rawabs_1demo_s10_224_lora_base_h10_2k`、`..._aefull_h10_2k`、`..._aefull_h10_10k_lr5` 和脚本 `scripts/pi05/run_pi05_native8_rawabs_1demo_s10_aefull_h10_10k_lr5.sh`。`aefull_h10_10k_lr5` 从本地 `pi05_base/params` 启动，只训练 action expert/action heads/time MLP，`discrete_state_input=False`、horizon=10、LR `5e-5`、EMA off；checkpoint `/data/Data4TB/01Proj/13atec/openpi_atec_runs/checkpoints/pi05_atec_task_e_native8_rawabs_1demo_s10_224_aefull_h10_10k_lr5/atec_task_e_pi05_native8_rawabs_1demo_s10_224_aefull_h10_10k_lr5_202606091220_aefull_h10_10k_lr5/9999` 约 `8.8G`。离线 probe 从 2k 的 mean first/chunk L2 `1.51/1.65` 改善到 10k 的 `0.28/1.10`；固定 zero noise 约 `0.22/0.87`，但 chunk 仍不稳。Isaac seed11 实测：`CHUNK_EXEC_STEPS=1` 得 `0.00`（视频 `logs/videos/task_e_pi05_native8_eval/pi05_native8_pi05_atec_task_e_native8_rawabs_1demo_s10_224_aefull_h10_10k_lr5_seed11_20260609122037.mp4`），`CHUNK_EXEC_STEPS=10` 得 `0.00`（视频 `..._20260609122341.mp4`，montage `..._20260609122341.montage.jpg`），`CHUNK_EXEC_STEPS=10 + ATEC_PI05_ZERO_NOISE=1` 仍 `0.00`（日志 `logs/pi05_native8_eval_20260609122748`）。结论：OpenPI server、unnormalize、env action scale 都已被验证会执行，但策略在 home/高位观测附近缺少连续相位，动作在 home 与中段轨迹之间抖动，不能形成有效 pick/place；该 pi0.5 rawabs 分支仍不得作为最后一次官网提交候选。
  - pi0.5 数据策略，2026-06-08/09：先不追加 GraspGen/PCA 数据。理由：旧 pi0.5 的主要疑点是 embodiment/action-time bridge，若 native8 视频完全不动，加数据只会放大错误；20-demo native8 s2 已能得到 seed11 `6`，因此当前先跑 100-demo stride=2 门控，比 stride=1 快且保留抓取时序。若 100-demo native8 s2 仍低于 15，才追加数据：LoRA/小 adapter 目标 `50-150` 条高质量真实观测-动作 demo；全量/continued fine-tune 目标 `300-500` 条，覆盖 object_1/2/3、不同初始位置、不同抓取顺序。GraspGen/PCA 只能作为数据生成 oracle/候选器，最终训练数据必须是比赛观测和合法 8D env action，不能依赖 oracle object pose 做提交推理。
  - pi0.5 评估工具：`demo/solution_pi05.py` 是 websocket bridge；`scripts/pi05/eval_task_e_pi05.py` 是 Isaac 评估入口；`scripts/pi05/run_pi05_eval_checkpoint.sh <checkpoint>` 会启动 `scripts/serve_policy.py`，跑 seeds `11/12/13`，seed11 录视频；`scripts/pi05/summarize_pi05_eval.py` 汇总分数；`scripts/pi05/compare_pi05_vs_act_baseline.py` 对比 XSA 基线并打印 `HOLD_ACT_XSA` 或 `REVIEW_FOR_DEPLOY`。100-demo 分支已经自动输出 `HOLD_ACT_XSA`。
  - AnyGrasp 下一步建议：AnyGrasp/AnyGrasp SDK 值得作为短线工程方案尝试，但它不是直接替代整套策略的“100% 一键方案”。推荐接法是：用 SAM3/任务先验 mask 限定目标点云区域，用 AnyGrasp 生成 6D grasp pose 和 collision-free candidate，再接 ATEC Piper 的 IK/轨迹规划、真实 finger center 接触、短稳定 basket-center release。优先验证 object_3，然后 object_1、object_2；严禁用 oracle object pose/mask 做最终提交。需要额外核查 AnyGrasp license、模型权重来源和比赛复现审查材料。
  - AnyGrasp 配置尝试，2026-05-30：官方 SDK 已克隆到 `third_party/anygrasp_sdk`；Python 3.11 二进制已放置为 `third_party/anygrasp_sdk/grasp_detection/gsnet.so` 和 `third_party/anygrasp_sdk/grasp_detection/lib_cxx.so`。Isaac/训练 Python 是 `/home/ubuntu/envs/genmanip-isaac5-py311/bin/python`，环境观测为 Python `3.11.15`、torch `2.7.0+cu128`、CUDA runtime `12.8`、GPU `NVIDIA RTX PRO 6000 Blackwell Workstation`，`open3d==0.19.0`、`graspnetAPI==1.2.11` 可用，`pointnet2._ext` 需先 `import torch` 后再导入。SDK 依赖 `libcrypto.so.1.1`，当前用 Isaac ROS2 bridge 自带 OpenSSL 1.1 库临时解决：`/home/ubuntu/projects/manipdojo2026/micromamba/envs/genmanip-sim/lib/python3.10/site-packages/isaacsim/exts/omni.isaac.ros2_bridge/humble/lib`，运行时需把它放进 `LD_LIBRARY_PATH`。
  - AnyGrasp license/权重历史阻塞，2026-05-30，已在 2026-06-03 解除：官方 SDK 是授权二进制，`license_checker -f` 需要 `ifconfig`；本机没有系统 `ifconfig`，已加兼容包装 `tools/anygrasp/ifconfig`。当前机器 feature id 是 `9024425463046318996`。2026-05-30 时尚未拿到 `licenseCfg.json` 和 `checkpoint_detection.tar`，当时不能宣称 AnyGrasp 已能跑 ATEC 抓取；2026-06-03 已收到并安装 license/权重，见下方 smoke 记录。
  - AnyGrasp MinkowskiEngine 已安装，2026-05-30：官方 modified `MinkowskiEngine` 已克隆到 `third_party/anygrasp_sdk/dependencies/MinkowskiEngine`，分支 `cuda-12-1`。系统只有 GCC/G++ `13.3.0`，会触发 CUDA 12.8 的 `std::__to_address` 歧义；已新建隔离编译器前缀 `/home/ubuntu/envs/anygrasp-gcc11`，内含 conda-forge GCC/G++ `11.4.0`，并只 patch 该前缀的私有头文件 `/home/ubuntu/envs/anygrasp-gcc11/lib/gcc/x86_64-conda-linux-gnu/11.4.0/include/c++/bits/shared_ptr_base.h`，把两处 `auto __raw = __to_address(__r.get());` 改成 `auto __raw = std::__to_address(__r.get());`，没有改系统 `/usr/include`。系统缺 BLAS devel symlink，已建项目内链接 `third_party/anygrasp_sdk/dependencies/blas_lib/libblas.so -> /usr/lib/x86_64-linux-gnu/blas/libblas.so.3`，并用本地 `third_party/anygrasp_sdk/dependencies/blas_include/cblas.h` 头文件。成功编译命令要点：`CUDA_HOME=/usr/local/cuda-12.8`、`TORCH_CUDA_ARCH_LIST=12.0`、`MAX_JOBS=4`、`CC/CXX/CUDAHOSTCXX` 指到 `/home/ubuntu/envs/anygrasp-gcc11/bin/x86_64-conda-linux-gnu-gcc/g++`，`--blas_include_dirs=third_party/anygrasp_sdk/dependencies/blas_include`，`--blas_library_dirs=third_party/anygrasp_sdk/dependencies/blas_lib`。安装后 `/home/ubuntu/envs/genmanip-isaac5-py311/bin/python` 可导入 `MinkowskiEngine 0.5.4`。
  - AnyGrasp 本地检查入口：`tools/anygrasp/check_anygrasp_env.sh` 会设置 `PATH`/`LD_LIBRARY_PATH`，检查 SDK 文件、torch/CUDA、Open3D、GraspNetAPI、pointnet2、MinkowskiEngine、feature id 和 checkpoint 缺口。2026-05-30 复查结果：`gsnet.so`、`lib_cxx.so`、`open3d 0.19.0`、`graspnetAPI 1.2.11`、`pointnet2._ext`、`MinkowskiEngine 0.5.4` 均 OK；仍缺官方 `licenseCfg.json` 和 `checkpoint_detection.tar`/AnyGrasp 权重。
  - AnyGrasp license/权重安装脚本，2026-05-30：拿到官方 license zip 和 `checkpoint_detection.tar` 后，运行 `tools/anygrasp/install_license_and_checkpoint.sh <license_zip_or_dir> <checkpoint_detection.tar>`；它会把 license 安装到 `third_party/anygrasp_sdk/grasp_detection/license` 和 `third_party/anygrasp_sdk/grasp_tracking/license`，把权重放到 `third_party/anygrasp_sdk/grasp_detection/log/checkpoint_detection.tar`，并调用 `license_checker -c` 校验。安装后用 `tools/anygrasp/run_detection_demo.sh` 跑官方 example data detection demo；该脚本当前已验证会在缺 license 时快速报错，不会误启动半截流程。
  - AnyGrasp license/权重已到位并跑通，2026-06-03：用户从 Chenxi Wang 收到 AnyGrasp SDK license 邮件，附件位于 `/home/ubuntu/Downloads/license_KeweiChen.zip`；detection 权重从官方 Google Drive 下载到 `/home/ubuntu/Downloads/anygrasp_weights/checkpoint_detection.tar`，大小约 `283M`。已运行 `tools/anygrasp/install_license_and_checkpoint.sh /home/ubuntu/Downloads/license_KeweiChen.zip /home/ubuntu/Downloads/anygrasp_weights/checkpoint_detection.tar`，安装到 `third_party/anygrasp_sdk/grasp_detection/license`、`third_party/anygrasp_sdk/grasp_tracking/license` 和 `third_party/anygrasp_sdk/grasp_detection/log/checkpoint_detection.tar`。`license_checker` 打印 `KeweiChen.lic check passed`，虽然检查命令返回码偶发非 0，但 SDK 真实加载时 `license passed: True, state: FvrLicenseState.PASSED`。
  - AnyGrasp detection smoke 结果，2026-06-03：修复 `pointnet2` 安装状态后，`/home/ubuntu/envs/genmanip-isaac5-py311/bin/python` 可导入 `pointnet2._ext`、`pointnet2.pointnet2_utils`、`pointnet2.pointnet2_modules`。运行 `tools/anygrasp/run_detection_demo.sh` 成功完成官方 `grasp_detection/example_data` 推理，输出 20 个 grasp scores，最高分约 `0.47641015`。这证明 AnyGrasp SDK detection 已在本机可用；下一步可以写 ATEC bridge，从 Isaac RGB-D/点云裁剪出目标区域，调用 AnyGrasp 生成 grasp pose，再接 Piper IK/执行 primitive。
  - AnyGrasp ATEC bridge 已接入，2026-06-03：新增 `scripts/graspnet_task_e/anygrasp_adapter.py`，`scripts/graspnet_task_e/run_graspnet_pick.py` 支持 `--grasp_provider anygrasp`，封装入口是 `scripts/graspnet_task_e/run_anygrasp_pick.sh <object> <seed>`。adapter 在 Isaac Python 进程内预加载 ROS2 bridge 自带 `libcrypto.so.1.1` / `libssl.so.1.1`，并把 `tools/anygrasp/ifconfig` 注入 `PATH`，否则 SDK 会出现 OpenSSL 或 machine-id/license 检查问题。检测器纯加载 smoke：`/home/ubuntu/envs/genmanip-isaac5-py311/bin/python -c 'from scripts.graspnet_task_e.anygrasp_adapter import _load_anygrasp_detector; _load_anygrasp_detector()'`，应看到 `license passed: True`。
  - AnyGrasp Task-E smoke 结果，2026-06-03：object_3 在 `--use_task_quat --tcp_z_offset 0.040 --close_z_offset -0.020` 下成功，`inside=True`、`z_gain_lift=0.137`，视频 `logs/videos/task_e_anygrasp/anygrasp_obj3_seed11_taskquat_lowclose.mp4`；object_2 同参数成功，`inside=True`，视频 `logs/videos/task_e_anygrasp/anygrasp_obj2_seed11_taskquat_lowclose.mp4`。object_1 当前失败两轮：`anygrasp_obj1_seed11_taskquat_lowclose_retry.mp4` 和 `anygrasp_obj1_seed11_defaultquat_nooffset_lowclose.mp4`，共同现象是 SDK 只给低分/极窄 width 候选（约 `0.006-0.008m`），夹爪闭合但物体没有随 lift 上升。不要把 AnyGrasp 当前 bridge 误当作三物体满分策略；它已完成接入验证，但 object_1 仍需沿既有 Task-E 校准成功模板继续调。
  - AnyGrasp 使用建议，2026-06-03：当前最稳用法是把 AnyGrasp 作为 RGB-D grasp candidate/center/yaw prior，再强制接 Task-E 已验证执行参数（`--use_task_quat`、低 close z、finger-center/staged transport、basket stable release）。不要直接使用 AnyGrasp raw 6D wrist pose；object_3 raw pose 曾失败，换 Task-E quaternion 后才成功。object_1 后续优先复用旧 GraspNet 成功证据里的控制模板：upper-surface median 执行中心、加 `OBJ_GRASP_CENTER_OFFSETS[1]`、`--close_z_offset -0.020`、keep grasp quat through transport/place/open、长 transport 和真实 finger-center servo。
  - GraspGen/AgileX 第三方路线复查，2026-06-03：联网查到 AgileX 社区/教程推荐 PiPER 的 GraspGen 路线，`AgilexRobotics/GraspGen` 链接当前不可访问，但可访问镜像/作者仓库 `vanstrong12138/GraspGen` 有 `ros2_jazzy_version` 分支；已浅克隆到 `third_party/GraspGen`。AgileX 官方教学仓库 `agilexrobotics/Agilex-College` 已浅克隆到 `third_party/Agilex-College`，其 `piper/GraspGen` README 描述了 PiPER + YOLOv8-Pose/TensorRT + PCA/AABB + IK/夹爪控制的流程。核心代码在 `third_party/GraspGen/sam3/realsense-sam.py::compute_grasp_from_pca_aabb()`：3D PCA -> 主轴系 AABB -> 最短边作为夹持方向。
  - GraspGen-style PCA/AABB ATEC 接入，2026-06-03：新增 `scripts/graspnet_task_e/pca_aabb_adapter.py`，`run_graspnet_pick.py` 支持 `--grasp_provider pca`。object_1 smoke 命令使用 `--grasp_provider pca --object 1 --seed 11 --mask_provider oracle --use_task_quat --tcp_z_offset 0.040 --close_z_offset -0.020 --close_steps 160 --move_steps 200 --transport_steps 1400 --place_steps 260`，输出 `[PCA_AABB] points=5231 extents=(0.140,0.122,0.056) axis=2 width=0.056`，说明几何法给出的可夹宽度合理，明显不同于 AnyGrasp 的 `0.006-0.008m` 异常窄 width。但该次仍失败：`pick=(0.950,0.261)`、`z_gain_lift=-0.024`、`inside=False`，视频 `logs/videos/task_e_pca/pca_obj1_seed11_taskquat_lowclose.mp4`。原因判断：单视角 RGB-D/mask 只看到 object_1 可见侧面，PCA/upper-surface median 的执行中心偏向左侧面；需要结合已知物体尺寸/多视角/learned completion/旧 GraspNet 成功中心补全到真实夹持中心。
  - object_1 方块“戳侧面、不含入夹爪”修复，2026-06-03：用户根据视频指出方块在左侧位置时 Piper 夹爪只是顶/戳方块，没有让方块进入两指之间再闭合。复测确认：PCA/AABB 加 `--preclose_insert_dx 0.040 --close_z_offset -0.020` 仍失败，`z_gain_lift=-0.025`，trace 里 close 阶段 `min_finger_dist≈0.058m`，说明指尖中心离目标太远；失败视频 `logs/videos/task_e_pca/pca_obj1_seed11_insert_dx004.mp4`。真正有效修复是不要手动压低 close height，而是对齐 Task-E 采集器 object_1 参数：`OBJ_GRASP_CENTER_OFFSETS[1]=(0.025,0,0)`、`OBJ_GRASP_Z_OFFSETS[1]=0.072`、`OBJ_CLOSE_Z_OFFSETS[1]=0.020`，并使用实际 `link7/link8` finger-center servo。成功命令：`OMNI_KIT_ACCEPT_EULA=YES PYTHONUNBUFFERED=1 /home/ubuntu/envs/genmanip-isaac5-py311/bin/python scripts/graspnet_task_e/run_graspnet_pick.py --grasp_provider pca --object 1 --seed 11 --mask_provider oracle --use_task_quat --tcp_z_offset 0.072 --close_steps 180 --move_steps 200 --transport_steps 1400 --place_steps 260 --video_path logs/videos/task_e_pca/pca_obj1_seed11_cfgclose.mp4 --save_debug_npz logs/pca_task_e/obj1_seed11_cfgclose_debug.npz --headless`；结果 `inside=True`、`z_gain_lift=0.144`，trace reach `min_finger_dist≈0.006m`、close `≈0.024m`。结论：object_1 失败不是“夹爪力不够”或“必须换 AnyGrasp”，而是 TCP/指尖中心/闭合高度标定错误导致的预闭合碰撞。`run_graspnet_pick.py` 已把 `preclose_insert` stats 纳入 `[TRACE]`，后续若继续扫参数，必须看 reach/insert/close/lift 的 finger-center 误差再判断。
  - `demo/solution_pca.py` object_1 submit-style 修复进展，2026-06-04：不要覆盖 `demo/solution.py`/ACT 基线，`solution_pca.py` 仍是实验分支。最新单物体 seed11/12/13 均已成功入篮：seed11 曾输出 `score=6.00`、initial `(1.0945,0.2979,0.8833)`、final `(1.0582,-0.3256,0.8820)`；最新 tight 参数下 seed12 `score=6.00`、final `(1.0338,-0.2790,0.8584)`，seed13 `score=6.00`、final `(1.0257,-0.2567,0.8871)`。验证入口：`ATEC_PCA_OBJECTS=1 OMNI_KIT_ACCEPT_EULA=YES PYTHONUNBUFFERED=1 /home/ubuntu/envs/genmanip-isaac5-py311/bin/python scripts/graspnet_task_e/debug_solution_pca_execution.py --seed <11|12|13> --object 1 --max_steps 4800 --headless`。关键参数：object_1 禁用 ROI fill-hole 点云中心；可见面中心补全 `OBJ_CENTER_COMPLETION_OFFSETS[1]=(0.020,0.0)`；`OBJ_GRASP_CENTER_OFFSETS[1]=(0,0)`；reach 阶段不要用低位 finger-z servo；object_1 release/open 高度用 runner 标定 `TABLE_TOP_Z+0.32/+0.24`；`OBJ1_HOLD_GAP=0.0415`，`OBJ_LIFT_STEPS[1]=300`，close/lift/mid 首次到 gap 阈值后记录 `_obj1_hold_grip`，后续 lift/mid/release/settle 保持该夹爪开度，避免闭到 0 把方块挤飞。错误经验：`OBJ_PLACE_XY_OFFSETS[1]=(0,-0.250)` 会让 release 目标变成 y=-0.55，导致物体掉在篮子前；object_1 rescue 的宽 ROI/白色网格容易误检，默认应关闭，除非重新做可靠前景分割。
  - `demo/solution_pca.py` object_2/3 状态，2026-06-04：object_2/3 仍未达到可提交稳定性，不要宣称 PCA 方案已满分。runner 对照中 object_3 seed11 曾成功一次，但 2026-06-04 复测同类 runner 参数也失败，说明旧成功不是稳态。5 月 20 日 `obj321_servo` 采集器日志证明官方 state_machine/真实 object root 下三物体可满分，object_3 成功 trace 常见 `z_gain≈0.42`、`gap_close≈0.063-0.069`、`finger_close≈0.016-0.030`、`finger_vec_close.x≈-0.01~-0.03`、`basket_inside=True`。submit-style object_3 关键新进展：低位/高位默认均不稳，问题从“完全不接触”推进到“可强抬但运输滑落”；当前实验默认值调为 `OBJ_CLOSE_Z_OFFSETS[3]=0.030`、`OBJ3_HOLD_GAP=0.055`、`OBJ_FINGER_XY_OFFSETS[3].x=-0.030`、`OBJ_LIFT_STEPS[3]=40`，能在部分 seed11 placement 产生强抬升，例如 `logs/pca_solution_obj3_fxneg003_full_seed11_20260604.log` 输出 `score=3.00`、`max_z_gain=0.195`，短 lift 日志 `logs/pca_solution_obj3_shortlift_seed11_20260604.log` 输出 `max_z_gain=0.236`；但 mid/release 阶段仍滑回桌面，未入篮。失败边界：`hold_gap=0/0.045/0.060/0.065`、`close_z=-0.020/0/0.020/0.050/0.070/0.090`、PCA yaw、`ATEC_PCA_USE_FULL_IK=1`、`ATEC_PCA_OBJ3_OBJECT_SERVO=1` 均未稳定解决。下一步不要再质疑是否“完全没夹到”；应聚焦 object_3 运输保持/低位转移：抓到后立即横向、减少纯竖直 lift、或做 closed-drag/under-scoop 入篮。object_2 仍需单独重新对齐，不要套 object_1/object_3 的补偿。
  - `demo/solution_pca.py` object_3 滑落修复追加，2026-06-04 18:20 CST：按用户方案已在实验分支实现“短抬升 + 动态视觉 object-servo + 低位 fallback”方向，但 seed11 仍未过，不要部署。当前文件改动包括：`OBJ3_HOLD_GAP=0.066`、`OBJ3_HOLD_GRIP_DEFAULT=(0.032,-0.034)`、`OBJ3_CLOSE_MIN_STEPS=160`、`OBJ3_LOW_HOLD_STEPS=140`、`OBJ3_APPROACH_FINGER_Z=TABLE_TOP_Z+0.090`、`OBJ3_CLOSE_FINGER_Z=TABLE_TOP_Z+0.010`、`OBJ3_LIFT_FINGER_Z=TABLE_TOP_Z+0.080`、`OBJ3_OBJECT_SERVO_GAIN=0.12`、`OBJ3_OBJECT_SERVO_MAX_XY=0.040`、`OBJ3_FINGER_SERVO_MAX_XY=0.080`，并把 object_3 fallback 改成更早在 `obj3_mid` 停滞时触发，`OBJ3_FALLBACK_DRAG_Z=TABLE_TOP_Z+0.035`、fallback steps `60/260/260/100`。注意：不要 freeze 实际 qpos 作为 object_3 hold grip，之前会锁住非对称夹爪开度如 `(0.035,-0.026)`；应使用标定默认 hold grip。不要只靠 `gripper_base` z，object_3 reach/close/lift 需要显式 finger-center z。
  - object_3 最新 trace 边界，2026-06-04：`logs/pca_solution_fix/obj3_seed11_runner_vec_match_debug_20260604_181342.log` 证明旧默认 `OBJ_FINGER_Y_OFFSET=+0.003` 会让 close finger-center 落在香蕉中心 y 侧约 `+0.024m`，把香蕉侧向挤开；已把默认 y offset 改成 `-0.020`。`logs/pca_solution_fix/obj3_seed11_yoffset_neg020_debug_20260604_181533.log` 证明 y 回正，但 x 可能过负；临时覆盖 `ATEC_PCA_OBJ3_FINGER_X_OFFSET=-0.004 ATEC_PCA_OBJ3_FINGER_Y_OFFSET=-0.020` 的 `logs/pca_solution_fix/obj3_seed11_xyoffset_xneg004_yneg020_debug_20260604_181706.log` 得到较接近 runner 的 close 向量：实际 finger 约 `(x=+0.002,y=-0.015,z=-0.014)` 相对当前视觉中心，但仍 `score=0`，说明问题已从“finger 没到位”推进到“闭合后没有形成可携带接触/视觉中心可能被夹爪遮挡干扰”。低位早触发 fallback `logs/pca_solution_fix/obj3_seed11_early_low_fallback_20260604_181908.log` 只把 object_3 最终推到 `(1.083,0.055,0.841)`，x 接近篮筐但 y 仍离篮筐中心 `+0.355m`；当前 closed-drag 高度/接触方向仍推不动香蕉。
  - object_3 下一步建议，2026-06-04：继续主攻 submit-style 规划抓取，不要回退 ACT/训练，也不要改官方物理。优先实现并验证真正的 table-level under-scoop/sweep：从当前视觉中心的 `+y` 一侧或弧线外侧进入，finger z 再低于 `TABLE_TOP_Z+0.035` 小步扫向 `BASKET_CENTER_Y`，同时保留 object_3 当前中心重估；fallback 要在 `obj3_mid` 更早触发并保证 1500 step 内完成。若继续调夹持，目标 trace 仍以 runner 成功 `logs/graspnet_task_e_obj3_sam3_closed_drag.log` 为准：`inside=True`、`z_gain_lift=0.207`、reach/close/lift `min_finger_vec≈(-0.007~-0.010, -0.002~+0.002, -0.014~-0.011)`、`gap≈0.066-0.068`。任何新参数必须用 `ATEC_PCA_DEBUG_TARGET=1` 或 `debug_solution_pca_execution.py` 看 finger-center/object-center 数值，不要只看视频判断。
  - object_3 用户指定“短抬升 + 低位 cradle/servo + 篮筐稳定释放”实现记录，2026-06-04 20:35 CST：已在 `demo/solution_pca.py` 实验分支实现，但仍未达到可部署。关键改动：`OBJ3_HOLD_GAP=0.0675`、`OBJ3_CLOSE_STEPS=90`、`OBJ3_LIFT_STEPS=35`、`OBJ3_LOW_HOLD_STEPS=0`、`OBJ3_LIFT_FINGER_Z=TABLE_TOP_Z+0.045`、`OBJ3_FINGER_X_OFFSET=-0.010`；hold grip 改为只在 `obj3_lift` 且 gap 达阈值后 latch，避免 close 阶段过早锁住；`obj3_basket_hold` 修掉一个明确 bug：不要把 `dynamic_finger_xy` 设成当前香蕉中心 `c_now`，应保持 `object_target_xy + carry_offset`，否则手会被拉回旧位置。验证日志：`logs/pca_solution_fix/obj3_seed11_eval_close90_shortlift_fasttransport_20260604_201931.log` 显示 latch 成功但 basket_hold desired 错误，最终 object_3 仍在 `y≈0.068`；修复后 `logs/pca_solution_fix/obj3_seed11_eval_fixservo_transport440_20260604_202219.log` 和 `...fingerik_20260604_202355.log` 仍卡在 `finger y≈-0.09`；低位 closed-drag 版本 `logs/pca_solution_fix/obj3_seed11_eval_low_closed_drag_20260604_202611.log` 也只到 `finger y≈-0.09`，object 未进篮；position-only DLS `logs/pca_solution_fix/obj3_seed11_eval_lowdrag_positiononly_*.log` 可让 gripper_base 更靠近篮筐，但 finger z 飘高到 `≈1.0m`，接触丢失。结论：当前 object_3 抓取位姿/短抬升已经可做到，主 blocker 是 submit-side Pinocchio DLS/运输控制没有复现 runner/collector 的 Isaac `CartesianController` 真实 PhysX Jacobian 行为，尤其是篮筐方向 y 运动和 finger z 保持；不要再把首要问题归咎为 GraspNet/AnyGrasp 候选模型。
  - object_3 后续最短路径，2026-06-04 20:35 CST：优先复刻 IsaacLab `CartesianController` 的动作生成细节，而不是继续换 grasp 模型。已确认官方 `get_action_spec()` 只允许改 action group 的 `mode/scale/clip`，不能直接提交 Cartesian action；若用它，只能尝试 arm velocity/scale 诊断，不能根治位姿控制。下一步建议：1) 用 `scripts/graspnet_task_e/debug_solution_pca_execution.py` 增加卡住阶段 q6/关节限位打印，确认是 joint limit 还是 DLS/rotation error；2) 将 `demo/solution_pca.py` 的 `_rot_error`/DLS 阻尼/每步 clamp 对齐 `source/atec_rl_lab/atec_rl_lab/utils/cartesian_controller.py`，优先用 position+current wrist quat 而不是强制 topdown；3) 若仍不行，做一个 small joint-space transport library：由 runner 在若干 object_3 y-band 初始位置生成相对 joint waypoints，submit 通过视觉中心选择/插值，而不是每步在线求 Cartesian IK。保持 `demo/solution.py` / `demo/policy_act.pt` ACT/XSA 基线不动，`solution_pca.py` 只作实验分支。
  - PCA/AABB + RGB-D band mask 三物体单体 smoke，2026-06-03：目标是避开 AnyGrasp SDK 的不可分发限制，形成可复现的几何抓取替代路线。`scripts/act/task_e/config.py` 的 `OBJ_SPAWN_Y_BANDS` 已对齐官方 `source/atec_rl_lab/atec_rl_lab/tasks/task_e/env_cfg.py`：object_1 `[+0.25,+0.29]`、object_2 `[+0.14,+0.20]`、object_3 `[+0.03,+0.09]`；`scripts/graspnet_task_e/tuntun_adapter.py` 的 `rgbd_band_object_mask()` 加了 object-specific z gate，避免 table/arm/篮子点混入。object_1 band 成功命令：`... run_graspnet_pick.py --grasp_provider pca --object 1 --seed 11 --mask_provider band --use_task_quat --tcp_z_offset 0.072 --close_steps 180 --move_steps 200 --transport_steps 1400 --place_steps 260 --video_path logs/videos/task_e_pca/pca_obj1_seed11_band_cfgclose_yband_zfilter.mp4 ...`，结果 `inside=True`、`z_gain_lift=0.144`。object_2 在官方 y-band 收窄后复测通过：`... --object 2 --seed 11 --mask_provider band --use_task_quat --tcp_z_offset 0.040 --close_z_offset -0.020 --transport_steps 1400 --place_steps 260 --video_path logs/videos/task_e_pca/pca_obj2_seed11_band_yalign_taskquat_lowclose.mp4 ...`，结果 `inside=True`、`z_gain_lift=0.119`。object_3 初始 band/PCA 能抓起但高空 release 会滑出；修复是在 `scripts/graspnet_task_e/pca_aabb_adapter.py` 中对 `object_index==3` 使用 oriented AABB center 而非 upper-surface median，并在 `run_graspnet_pick.py` 增加 `--open_release_z` 支持“高位运输、低位开爪”。object_3 seed11 AABB-center 高位 release 可成功一次：`logs/videos/task_e_pca/pca_obj3_seed11_band_aabbcenter_taskquat_lowclose.mp4`，结果 `inside=True`、`z_gain_lift=0.161`；seed12 高位 release 会滑出，改用 `--release_z 0.32 --open_release_z 0.24 --place_steps 220` 后成功，视频 `logs/videos/task_e_pca/pca_obj3_seed12_band_aabbcenter_highcarry_lowopen.mp4`，结果 `inside=True`、最终篮内 delta 约 `(-0.031,-0.049)`。注意：这些是单物体 runner 验证，仍需串成三物体连续 episode 并迁移到 `demo/solution.py` 的 observation-only 控制逻辑；当前 `run_graspnet_pick.py` 仍会读取 env object state 做验证/动态 servo，不可直接当作最终提交。
  - Submit-style PCA controller debug，2026-06-04：`demo/solution_pca.py` 是实验分支，未部署；`demo/solution.py` 仍是 ACT/XSA baseline。`scripts/act/eval_task_e_act.py` 已支持 `--solution_module solution_pca`，`scripts/graspnet_task_e/debug_solution_pca_execution.py` 可用 Isaac internals 量化 `finger_dist/max_z_gain/pin_mean_err`。已确认 Pinocchio FK 与 Isaac link7/link8/gripper_base 对齐，`pin_mean_err finger=0.0000 gb=0.0000`，所以“根本没去夹/没夹起”的主因不是 FK 坐标系错误，而是 submit-style IK/控制器没有复现 `run_graspnet_pick.py` 的 Isaac `CartesianController` 低位接触能力。
  - Submit-style object_1 当前状态，2026-06-04：已尝试把 `run_graspnet_pick.py` 的 object_1 成功模板迁移到 `demo/solution_pca.py`，包括 RGB-D world AABB 中心补全、`OBJ_CENTER_COMPLETION_OFFSETS[1]=(0,+0.018)`、object_1 x 执行 offset 从 `+0.025` 收到 `+0.010`、低 close plane `TABLE_TOP_Z+0.030`、以及 direct finger-center IK。结论：感知中心可做到厘米级，但当前 Pinocchio IK 在 close 阶段仍常把 link7/link8 finger-center 卡在物体上方约 `+0.05m`，`max_z_gain≈0`，object_1 未抬起。direct finger-center IK 和强制 object_1 task quaternion `(0,0.707,0.707,0)` 反而更差，已回到 gripper_base compensation 路线；不要误判为 Grasp/PCA/AnyGrasp 本身失败。下一步应复刻 Isaac `CartesianController` 的 differential IK 数学/参数，或在 submit policy 内实现更可靠的 DLS resolved-rate joint update，而不是继续换抓取模型。
  - Submit-style object_1 preclose insert 更新，2026-06-04：`demo/solution_pca.py` 新增 `_PiperIK.step_dls()`、runner-style finger-center XY correction、object_1 task quaternion `[0,0.709,0.705,0]`、object_1 relative-z correction、preclose insert 和 object_1 release overshoot。视频 `logs/videos/task_e_pca_solution/pca_solution_obj1_seed11_best_trace_20260604_114715.mp4` 明确显示无 insert 时夹爪嘴停在方块侧边，抬起时方块留在桌面。加入 insert 后，`OBJ_PRECLOSE_INSERT_OFFSETS[1]=(0,-0.040)` 曾得到 `score=3`、`max_z_gain=0.137m`，但运输中滑落到 `y≈-0.047`；`(0,+0.025)` 曾得到 `max_z_gain=0.193m`、最终 `y≈-0.083`，仍未进篮，另一次随机位置未抓起。结论：insert 方向/幅度是有效抓起的关键，但 object_1 仍未稳健；当前 `demo/solution_pca.py` 仍是 debug-only，不要覆盖 `demo/solution.py` 或 `demo/policy_act.pt`。
  - GraspNet object_1 对照，2026-06-03：尝试复跑旧已知成功命令，视频目标 `logs/videos/task_e_graspnet/graspnet_pick_obj1_seed11_recheck_20260603.mp4`，进程在 GraspNet 推理/候选生成阶段超过 3 分钟无 `[GRASP]` 输出，占 GPU 约 `6.6GB` 且 CPU 高负载，已手动 kill。该次不是抓取成功/失败证据；后续若要继续对比 GraspNet，先排查为什么当前 recheck 推理阶段变慢/卡住。
  - AnyGrasp 第三方参考结论，2026-05-30：未找到松灵/AgileX 官方 AnyGrasp for Piper 项目；AgileX 官方主要提供 PiPER Python/ROS/ROS2/URDF/仿真控制栈。第三方/论文路线支持“AnyGrasp 作为 pre-contact/contact-pose prior，再接机器人 IK/轨迹规划”：DKT 透明物体工作报告过 `RealSense D435 + PiPER Arm + AnyGrasp + CuRobo` 抓取栈；BayesVLA 把策略拆成视觉动作 prior 和语言 likelihood，并把 pre-contact 阶段交给 AnyGrasp 这类 frozen contact-pose model。对本项目的启发：AnyGrasp 应先做候选抓取姿态生成和 ranking，ATEC Piper 仍需要单独处理 TCP 标定、IK/trajectory、夹爪闭合/运输防滑、篮筐中心高位稳定 release。
  - 当前工作建议排序：1) 保持 `demo/policy_act.pt` 的 XSA-ACT 基线不动；2) 若追短期提分，优先把 AnyGrasp/SAM3 作为 grasp candidate generator 接入现有 calibrated Piper primitive，而不是继续盲训 pi0.5；3) 若继续 VLA，重做 OpenPI/Piper embodiment adapter（native Piper dims/cameras/norm/discrete_state_input），不要再使用 fixed-base 16D padding 当正式结论；4) 每次部署前必须独立进程评估 seeds `11/12/13`，至少保留一个视频和 `[BASKET]` 物体入篮日志。
- Latest status, 2026-05-20 20:55 CST:
  - Supersedes the older "object 1/2 unresolved" notes below; keep them only as historical debugging context.
  - L0 Task E three-object scripted data generation now works well enough to collect demos. Verified all-object order is `3 2 1` with `--only_success --trace --abort_failed_lift`.
  - Verified object-1-only success: `datasets/atec_task_e_probe_obj1_graspoffset_servo/trajectory.hdf5`, 3/3 successful demos; trace z-gain about `0.387-0.417`, `basket_inside=True`.
  - Verified object-2-only success: `datasets/atec_task_e_probe_obj2_original_rootservo/trajectory.hdf5`, 3/3 successful demos; trace z-gain about `0.302-0.304`, `basket_inside=True`.
  - Verified three-object success: `datasets/atec_task_e_probe_obj321_final_servo/trajectory.hdf5`, 3 successful demos in 4 attempts; success traces have object_1/2/3 `basket_inside=True`.
  - Review video: `logs/videos/task_e_obj321_final_servo/demo_0000.mp4`; keyframe summary: `logs/videos/task_e_obj321_final_servo/summary_5s.jpg`.
  - Root cause/fix: data generation was wrong, not ACT training. Object 1 must keep the finger-centre servo target on `obj_pos + OBJ_GRASP_CENTER_OFFSETS[1]`; object 2 must track the USD/root object position, not the offset center. Both need closed-loop finger-centre servo during `REACH/CLOSE/LIFT/TRANSPORT/PLACE`.
  - Stable constants in `scripts/act/task_e/config.py`: object 1 offset `(0.025,0,0)`, grasp z `0.072`, close z `0.020`, transport/place steps `2200/300`; object 2 offset `(0.060,0,0)`, grasp z `0.135`, transport/place steps `2200/300`; default `PICK_OBJECTS="3 2 1"`.
  - `scripts/act/task_e/collector.py` has the important per-object servo distinction: object 1 servo/trace target uses `obj_pos + OBJ_GRASP_CENTER_OFFSETS`; object 2/3 use `obj_pos`. Do not "simplify" this back to one formula.
  - `scripts/act/train_task_e.py` uses `lr_drop = max(int(2/3 * total_iters), 1)` so tiny smoke tests no longer fail with `StepLR(step_size=0)`.
  - CAP-X / RoboTwin lesson: use contact/TCP calibrated grasp candidates and closed-loop place logic first. GraspNet/ContactGraspNet can be added later as a candidate generator, but it is not required to fix this current oracle data bug.
  - Real pipeline command:
    `DATA_DIR="datasets/atec_task_e_obj321_servo_100demos" PICK_OBJECTS="3 2 1" NUM_DEMOS=100 MAX_ATTEMPTS=260 TOTAL_ITERS=100000 BATCH_SIZE=512 SEED=1 LOG_FREQ=100 SAVE_FREQ=5000 RUN_NAME="act-task-e-rgb-obj321-servo-100demos-seed1" bash scripts/act/run_task_e_pipeline.sh`.
  - Active long run started via systemd user unit because plain background/nohup could be killed with the Codex command: unit `atec-task-e-obj321-servo-20260520211650.service`, log `logs/task_e_obj321_servo_pipeline_20260520_211650.log`, latest symlink `logs/task_e_obj321_servo_pipeline.latest`.
  - As of 2026-05-20 21:22 CST the long run is active, `Demo 1/100` succeeded and wrote `datasets/atec_task_e_obj321_servo_100demos/trajectory.hdf5` (~1.3 GB), then proceeded to `Demo 2/100`.
- Latest train/eval status, 2026-05-21 14:55 CST:
  - `atec-task-e-obj321-servo-20260520211650.service` completed normally; raw dataset is `datasets/atec_task_e_obj321_servo_100demos/trajectory.hdf5` (~130 GB), filtered dataset is `trajectory_filtered.hdf5` (~35 GB), run is `runs/act-task-e-rgb-obj321-servo-100demos-seed1`.
  - Checkpoints include `best_loss.pt` and `final.pt`; training reached `100000` iters, final logged loss around `0.0100`.
  - First ACT eval of `best_loss.pt` scored `0/3` episodes; video `logs/videos/task_e_act_eval/obj321_best_loss_100demos_20260521_145104.mp4` shows the arm staying near the basket/home pose instead of starting the object grasp.
  - Root cause found in `demo/solution_act.py`: the home stage treated env action as PD/velocity, but Task E action is `(joint_target - default_joint_pos) / 0.5`. Fix is to command the static teleop-home target action until stable, then switch to ACT. Re-run eval after this fix before judging model quality.
  - After the home fix, `best_loss.pt` eval improved: 1-episode score `9.00` with video `logs/videos/task_e_act_eval/obj321_best_loss_homefix_1ep_20260521_145529.mp4`; same-process 3-episode eval had valid first episode score `18.00` and video `logs/videos/task_e_act_eval/obj321_best_loss_homefix_3ep_20260521_145652.mp4`, but subsequent episodes reset to `done=True` at step 1, so do not treat that mean score as independent.
  - `final.pt` single-episode eval scored `6.00`, so deploy `best_loss.pt`.
  - Deployment: previous `demo/policy_act.pt` backed up as `demo/policy_act.prev_20260521_145951.pt`; current `demo/policy_act.pt` is copied from `runs/act-task-e-rgb-obj321-servo-100demos-seed1/checkpoints/best_loss.pt` and sha256 matches `76bbef2c81c85fc6cc60448476ff09075ec8bc97217c2d2fb70a3253fd3e307d`.
- XSA ACT experiment, 2026-05-21:
  - Local XSA package source: `/home/ubuntu/Documents/01Proj/13atec/xsa_update_20260520/lerobot_policy_act_xsa.zip`; it is a LeRobot `act_xsa` plugin with 12-dim example config and is not directly runnable for ATEC Task E's 8-dim Piper setup.
  - Minimal XSA adaptation added behind `--use_xsa`: `source/atec_rl_lab/atec_rl_lab/train/act/act/detr/transformer.py` and `demo/act/detr/transformer.py` support `XSATransformerEncoderLayer` / `XSATransformerDecoderLayer`; default ACT remains unchanged.
  - `scripts/act/train_task_e.py` now saves `model_args` in checkpoints; `demo/solution_act.py` reads `model_args` including `use_xsa`, so ordinary ACT and XSA checkpoints can both load.
  - Smoke test passed: `runs/act-task-e-rgb-xsa-smoke-10demos-seed7/checkpoints/final.pt` loads as `XSATransformerEncoderLayer`.
  - Active XSA comparison run: systemd user unit `atec-task-e-xsa-30k-20260521165009.service`, log `logs/task_e_xsa_30k_20260521_165009.log`, run name `act-task-e-rgb-xsa-100demos-30k-seed11`, command uses `--use_xsa --total_iters 30000 --batch_size 256 --num_demos 100`.
  - Mid-training evals are noisy but not yet encouraging: 19k-ish snapshot scored `9/18` with video `logs/videos/task_e_act_eval/xsa_mid_best_1ep_20260521_174124.mp4`; later current-best snapshot at `20260521_175038` scored `0/18` with video `logs/videos/task_e_act_eval/xsa_current_best_1ep_20260521_175038.mp4`. Do not infer ranking from training loss alone; run independent single-episode processes because same-process multi-episode reset can be invalid after success.
- GraspNet / TunTunClaw integration, 2026-05-21:
  - User reference cloned into `third_party/tuntunclaw/` from Datawhale/OpenClaw home assistant material. Official GraspNet checkpoint is `third_party/tuntunclaw/temp/logs/log_rs/checkpoint-rs.tar`.
  - Local bridge files: `scripts/graspnet_task_e/tuntun_adapter.py` and smoke runner `scripts/graspnet_task_e/run_graspnet_pick.py`.
  - Dependencies were made to work in `/home/ubuntu/envs/genmanip-isaac5-py311/bin/python`: `open3d==0.19.0`, local `graspnetAPI`, `transforms3d==0.4.2`, and the GraspNet `pointnet2` CUDA extension built for Blackwell with `TORCH_CUDA_ARCH_LIST=12.0`.
  - Important lesson: raw GraspNet point/orientation alone did not lift object 1. The first successful object-1 smoke used GraspNet/TunTunClaw for candidate generation, but executed with Task-E calibrated controls: robust mask upper-surface median as execution center, object-1 `OBJ_GRASP_CENTER_OFFSETS` added, `--close_z_offset -0.020`, finger-center servo, keep grasp quaternion through transport/place/open, and long transport.
  - Known-good object-1 command:
    `python scripts/graspnet_task_e/run_graspnet_pick.py --object 1 --seed 11 --tcp_z_offset 0.040 --close_z_offset -0.020 --use_task_quat --close_steps 160 --move_steps 200 --transport_steps 1400 --place_steps 260 --headless --device cuda:0 --video_path logs/videos/task_e_graspnet/graspnet_pick_obj1_seed11_keepquat_slow.mp4 --save_debug_npz logs/graspnet_task_e/obj1_seed11_keepquat_slow_debug.npz`.
  - Verified result for that command: `inside=True`, `z_gain_lift=0.164`, final object_1 position about `(1.043,-0.247,0.934)`, basket delta `(-0.037,+0.053)`. Video: `logs/videos/task_e_graspnet/graspnet_pick_obj1_seed11_keepquat_slow.mp4`.
  - Object 3 GraspNet smoke also succeeded with refined AABB oracle mask and lower close height: command uses `--object 3 --tcp_z_offset 0.055 --close_z_offset 0.015 --use_task_quat --place_x_offset 0.04 --place_y_offset 0.05`; result `inside=True`, `z_gain_lift=0.188`, video `logs/videos/task_e_graspnet/graspnet_pick_obj3_seed11_lowclose.mp4`.
  - Object 2 GraspNet smoke is not solved yet. Best attempts can move/lift slightly (`z_gain_lift≈0.07-0.10`) but final y remains outside basket; simple `--post_push` moved it only to about `y=-0.056`. Keep using the existing Task-E calibrated state-machine/ACT path for object 2 until a better GraspNet contact primitive is found.
  - Added legal-ish RGB-D y-band mask provider `rgbd_band_object_mask()` and runner flag `--mask_provider band`; it avoids sim object state by using camera intrinsics, depth, table height, and official spawn y-bands. Current band-mask object-3 test did not lift yet, so the GraspNet success videos above still rely on `oracle_object_mask()`/AABB debug masks. Do not use oracle masks in final submission without replacement.
  - Three-object stable oracle/state-machine verification after GraspNet changes: `scripts/act/collect_demos_task_e.py --pick_objects 3 2 1 --num_demos 1 --max_attempts 3 --only_success --trace --abort_failed_lift --save_video ...` succeeded on attempt 2, with all object_1/2/3 `basket_inside=True`. Video: `logs/videos/task_e_obj321_current_verify/demo_0000.mp4`.
- SAM3 segmentation / basket-center release, 2026-05-21:
  - Local SAM3 checkpoint exists at `/home/ubuntu/Documents/01Proj/sam3d_gs/submodule/Prompt-Inpaint/checkpoints/sam3.pt`; SAM3 is installed in `/home/ubuntu/Documents/01Proj/sam3d_gs/.venv/bin/python`, not in the Isaac env. Keep SAM3 isolated as a subprocess rather than installing it into `/home/ubuntu/envs/genmanip-isaac5-py311`.
  - Added `scripts/graspnet_task_e/sam3_segment_image.py`: standalone RGB image -> SAM3 binary mask. It uses local Prompt-Inpaint `SAM3Predictor` and writes `.npy` mask plus optional JSON metadata.
  - `scripts/graspnet_task_e/run_graspnet_pick.py` now supports `--mask_provider sam3`, `--sam3_prompt`, `--sam3_threshold`, and `--sam3_python`. It also defaults to `--basket_center_release`: during transport/place/open it servo-corrects the EE target by current object center error so the object center is driven to `(BASKET_CENTER_X, BASKET_CENTER_Y)` before opening.
  - SAM3 smoke on saved frames passed for all three prompts: object 1 best prompt `box` score `0.863`, mask `31206` px; object 2 best prompt `yellow bottle` score `0.961`, mask `2509` px; object 3 best prompt `banana` score `0.953`, mask `1311` px. Overlays are `logs/graspnet_task_e/sam3_smoke_obj{1,2,3}_overlay.png`.
  - First live object-3 SAM3+GraspNet smoke was blocked before environment creation by `ModuleNotFoundError: scripts.act`; fixed by explicitly inserting repo root into `sys.path` at the top of `run_graspnet_pick.py`. Re-run live smoke after this fix.
  - Live object-3 SAM3+GraspNet with default high close failed (`z_gain_lift=0.000`). Lowering close offset fixed the contact: command used `--mask_provider sam3 --object 3 --seed 11 --tcp_z_offset 0.055 --close_z_offset -0.005 --use_task_quat --basket_center_release`; result `inside=True`, `z_gain_lift=0.211`, final `(0.995,-0.254,0.852)`, basket delta `(-0.085,+0.046)`. Video: `logs/videos/task_e_graspnet/graspnet_pick_obj3_sam3_basket_center_lowclose.mp4`.
  - `run_graspnet_pick.py` now has a local GraspNet-only default close offset for object 3 (`GRASPNET_CLOSE_Z_DEFAULTS[3] = -0.005`), without changing global Task-E scripted collector constants.
  - Debug finding after user insisted the SAM-center-to-basket rule is valid: object-1 initial SAM3 failure was not a SAM capability issue. `sam3_segment_image.py` returned multiple masks and `run_graspnet_pick.py` initially picked the highest-confidence distractor near the basket/table (`y≈-0.316`) instead of the sugar box (`y≈+0.261`). Fix: `sam3_segment_image.py` now can write all candidates, and `run_graspnet_pick.py` selects/refines the candidate with the most RGB-D points inside the target object's official spawn x/y band and table-height range.
  - After world-band candidate selection, object-1 SAM3 perception is correct (`selected_candidate=1 prompt=box`, mask `3327` px, GraspNet/pointcloud target about `(1.020,0.261,0.917)`). It partially lifts (`z_gain_lift=0.120`) but misses the basket at final y `-0.051`; stronger basket servo made grasping worse, so current blocker is grasp/contact stability plus transport slip, not segmentation.
  - Object-2 SAM3 perception is also correct (`band_ratio≈0.98-0.99`, mask about `2500-2700` px). With default offset it partially lifts (`z_gain_lift≈0.098-0.112`) but stops around y `-0.08..-0.10`; no-offset is worse and lower close height still misses. Current blocker is bottle grasp/contact stability and transport slip, not SAM3/SAM2.
  - Do not switch to SAM2 as the primary fix unless SAM3 latency becomes unacceptable. The failure evidence points to candidate filtering (fixed) and execution/contact tuning (still pending), while SAM3 masks are good once selected by world-band constraints.
  - User correctly identified early release: added explicit transport-end logging, short stable-hold gating, and "do not open unless stable" logic in `run_graspnet_pick.py`. New args include `--release_z`, `--basket_xy_tol`, `--basket_hold_steps`, `--basket_stable_steps`, `--basket_recovery_steps`, `--auto_table_push_on_slip`, and `--drag_recovery_steps`.
  - Important nuance: long high hold can cause slip/bounce. Object-3 high-hold test reached `min_xy_err≈0.055` but failed because it waited too long and then dropped/bounced. Use a short stable hold (`basket_stable_steps` about `6-12`, `basket_xy_tol` about `0.09`) rather than hundreds of hold steps.
  - Verified object-3 SAM3 short-stable release success after adding `TRANSPORT_END`: command used `--object 3 --mask_provider sam3 --close_z_offset -0.005 --release_z 0.18 --basket_xy_tol 0.09 --basket_hold_steps 40 --basket_stable_steps 6 --transport_steps 220`; result `inside=True`, `TRANSPORT_END xy_err=0.006`, final `(1.077,-0.313,0.852)`. Video: `logs/videos/task_e_graspnet/graspnet_pick_obj3_sam3_closed_drag.mp4`.
  - Object-1 and object-2 still fail under SAM3+GraspNet despite correct masks. At `TRANSPORT_END`, both have slipped/dropped before release: object-1 example final transport `(0.885,-0.056,0.872)`, `xy_err=0.313`; object-2 example `(0.978,0.034,0.882)`, `xy_err=0.349`. Closed-drag/table-push recovery as currently implemented does not move them, likely because Piper `gripper_base` low push pose is not actually contacting the object side. Next step should switch recovery to the known Task-E calibrated state-machine transport/contact primitive or redesign pusher using actual finger/link contact geometry, not just gripper_base target z.
  - Follow-up on 2026-05-21 night: user explicitly required the method to be competition-real, so `scripts/graspnet_task_e/run_graspnet_pick.py` was changed to use actual `link7/link8` finger centers for transport/recovery, not fake `gripper_base` contact assumptions. Added staged transport (`--staged_transport`, `--transport_servo_fraction`) so the arm first carries smoothly and only applies basket-center object servo in the final segment.
  - Verified SAM3+GraspNet object-3 with real finger/staged transport: command used `--object 3 --mask_provider sam3 --close_z_offset -0.005 --transport_steps 700 --transport_servo_fraction 0.20 --basket_servo_gain 0.75 --basket_servo_max_xy 0.18`; result `inside=True`, final about `(0.995,-0.220,0.853)`, video `logs/videos/task_e_graspnet/graspnet_pick_obj3_sam3_staged_transport.mp4`.
  - SAM3+GraspNet object-1/2 are still not robust enough for submission. Object-1 no-offset test can partially lift (`z_gain_lift≈0.130`) but slips near `y≈-0.055`; object-2 offset test can partially lift (`z_gain_lift≈0.120-0.137`) but slips around `y≈-0.10`. GraspNet native yaw/quat made object-2 worse. The blocker is real contact stability during transport, not SAM3 segmentation.
- ACT obj321 servo training/eval, 2026-05-21:
  - Dataset exists: `datasets/atec_task_e_obj321_servo_100demos/trajectory.hdf5` (~122G) and `trajectory_filtered.hdf5` (~33G).
  - Training completed: run `runs/act-task-e-rgb-obj321-servo-100demos-seed1`, checkpoints through `final.pt`; training log symlink `logs/task_e_obj321_servo_pipeline.latest`.
  - Default submission checkpoint is already installed: `demo/policy_act.pt` SHA256 matches `runs/act-task-e-rgb-obj321-servo-100demos-seed1/checkpoints/best_loss.pt` (`76bbef2c81c85fc6cc60448476ff09075ec8bc97217c2d2fb70a3253fd3e307d`).
  - Evaluation with real observations: `final.pt` single episode scored `6.00`, video `logs/videos/task_e_act_eval/obj321_servo_final_1ep_20260521_realobs.mp4`. `best_loss.pt` produced one successful `18.00` episode with video `logs/videos/task_e_act_eval/obj321_servo_bestloss_3ep_20260521_realobs.mp4`, but independent single-process repeats scored `6/12/12`; treat current expected score as roughly 10-12 with occasional 18, not yet robust.
  - Known eval gotcha: same-process multi-episode reset can produce bogus later episodes with `steps=1`; use separate Python processes for trustworthy episode estimates.
  - 2026-05-21 night inference tuning: `demo/solution_act.py` now supports env vars `ATEC_ACT_TEMPORAL_AGG=0`, `ATEC_ACT_PREFER_NEW=1`, and `ATEC_ACT_TEMPORAL_K=<float>` without changing default behavior. `scripts/act/eval_task_e_act.py` now supports `--seed` and prints `[BASKET]` object status at the end.
  - Fixed-seed quick checks with `best_loss.pt`: prefer-new (`ATEC_ACT_PREFER_NEW=1 ATEC_ACT_TEMPORAL_K=0.05`) scored `6/15/0` on seeds `11/12/13`; no-agg (`ATEC_ACT_TEMPORAL_AGG=0`) initially scored `6/9/15` without basket debug, but seed handling/debug rerun of seed13 scored `0` and showed all objects still outside basket. More controlled seed sweeps are needed before changing submission inference defaults.
  - Started ACT seed2 training from the existing filtered dataset, no recollection: systemd user unit `atec-task-e-act-seed2-20260521225511.service`, log `logs/task_e_act_seed2_20260521225511.log`, run name `act-task-e-rgb-obj321-servo-100demos-seed2`.
  - Seed2 mid-training evals: `5000.pt` on fixed seed 11 scored `0.00` (all objects outside); then-current `best_loss.pt` scored `6.00` with only object_3 inside. `10000.pt` on fixed seed 11 scored `0.00`. `15000.pt` on fixed seed 11 also scored `0.00`; video `logs/videos/task_e_act_eval/seed2_15000_seed11.mp4`. `20000.pt` on fixed seed 11 scored `6.00` with only object_3 inside; video `logs/videos/task_e_act_eval/seed2_20000_seed11.mp4`. Do not replace `demo/policy_act.pt` with seed2 before later checkpoints prove better. Training continued past `Iter 20000`, latest logged loss around `0.020`.
  - 2026-05-22 01:12 CST seed2 reached `30000.pt`; loss about `0.0158`, and `best_loss.pt` was updated. Attempted fixed-seed eval `logs/eval_task_e_seed2_30000_seed11.log`, but it is invalid: GPU was already at ~91/98 GB because unrelated Ray/FSDP processes used ~76 GB, Isaac/Warp hit CUDA OOM and fell back to software physics. The eval process was killed before completion; do not treat this as model success or failure. Next trustworthy seed2 eval should wait for enough free CUDA memory, or after the unrelated Ray job finishes; avoid stopping seed2 because `scripts/act/train_task_e.py` has no resume argument.
  - 2026-05-22 01:19 CST valid CUDA eval of seed2 `30000.pt` on fixed seed 11 scored `0.00`; all objects outside. Video: `logs/videos/task_e_act_eval/seed2_30000_seed11_valid.mp4`. The original seed2 service then failed with status 143 at iter ~30400, leaving `30000.pt` as the last full checkpoint.
  - Added resume support to `scripts/act/train_task_e.py` and `scripts/act/run_task_e_pipeline.sh`: `--resume_checkpoint`, `--resume_iter`, and env vars `RESUME_CHECKPOINT` / `RESUME_ITER`. New checkpoints save `train_iter`, optimizer, scheduler, and EMA state; legacy checkpoints can resume weights with a fresh optimizer. Smoke test passed: `logs/train_task_e_resume_smoke_seed2_30000.log`, run `runs/act-task-e-resume-smoke-seed2-30000`.
  - Started resumed seed2 continuation: systemd unit `atec-task-e-act-seed2-resume-20260522013429.service`, log `logs/task_e_act_seed2_resume_20260522013429.log`, command resumes from `runs/act-task-e-rgb-obj321-servo-100demos-seed2/checkpoints/30000.pt` with `RESUME_ITER=30000` and `TOTAL_ITERS=100000`.
  - Resume status 2026-05-22 01:52 CST: service successfully loaded `100` trajectories (`169700` timesteps), resumed at iter `30000`, and reached about `Iter 31000`; GPU utilization was 100%, process `722380` used about 13 GB VRAM. New `30000.pt` / `best_loss.pt` are about 333 MB because they now include optimizer/scheduler/EMA state.
  - Seed2 resumed eval: `35000.pt` exists and valid seed-11 eval scored `0.00` (`logs/eval_task_e_seed2_35000_seed11_valid.log`); all object_1/2/3 were outside. Do not deploy seed2 at 35k. Continue training toward 50k/100k and only evaluate later checkpoints or best_loss when GPU is not crowded.
  - Seed2 `best_loss` snapshot at 2026-05-22 02:22 (`runs/act-task-e-rgb-obj321-servo-100demos-seed2/checkpoints/best_loss_snapshot_20260522_022223.pt`, sha256 `5e52307e84684a98429273102e136f0398e458ab1bd02a0dc6efc025493d4f02`) scored `12.00` on seed 11 (`logs/eval_task_e_seed2_bestloss_snapshot_20260522_022223_seed11.log`): object_2 and object_3 inside, object_1 outside. This is better than `35000.pt` but still not enough to replace XSA; evaluate later best_loss/50k across multiple seeds before deployment.
  - Resume status 2026-05-22 02:32 CST: training reached about `Iter 41200`; `40000.pt` exists (`runs/act-task-e-rgb-obj321-servo-100demos-seed2/checkpoints/40000.pt`). No 40k eval yet because GPU was crowded by another Ray/FSDP job (~70 GB) plus seed2 training; wait for 50k or a clear GPU window.
  - Seed2 `best_loss` snapshot at 2026-05-22 02:35 (`runs/act-task-e-rgb-obj321-servo-100demos-seed2/checkpoints/best_loss_snapshot_20260522_023530.pt`, sha256 `4048a27189b6539e9c5a699ae39eb4e2ea9496e197649d5fa5434141b5495ad7`) scored only `6.00` on seed 11 (`logs/eval_task_e_seed2_bestloss_snapshot_20260522_023530_seed11.log`): only object_3 inside. Lower training loss did not correlate with better task score; do not deploy this snapshot.
  - Seed2 numbered checkpoint evals after resume: `45000.pt` scored seed11 `18.00` but seed12 `0.00`; `50000.pt` scored seed11 `15.00` and seed12 `6.00`. This is still too unstable to deploy over current XSA candidate. Continue training and evaluate later checkpoints; object_1 remains the main missing piece.
  - Resume status 2026-05-22 03:24 CST: `55000.pt` exists and `best_loss.pt` updated at the same time, but no 55k eval was run because 45k/50k already proved instability. Next planned evaluation gate is 75k or final/100k unless training fails.
  - Resume status 2026-05-22 04:00 CST: training reached about `Iter 64300`; `60000.pt` exists and `best_loss.pt` updated at about `03:58:59`. No 60k eval; continue to 75k gate.
  - Seed2 `75000.pt` eval on seed 11 scored only `6.00` (`logs/eval_task_e_seed2_75000_seed11.log`): only object_3 inside. Do not spend more seed12/13 eval on 75k; wait for final/100k or a specific promising snapshot.
  - Seed2 mid/late checks on 2026-05-22: `85000.pt` seed11 scored `0.00` (`logs/eval_task_e_seed2_85000_seed11.log`), while the then-current `best_loss.pt` around iter `87300` scored `15.00` (`logs/eval_task_e_seed2_bestloss_87300_seed11.log`) with object_2 and object_3 inside and object_1 outside. This reinforces that training loss and task score are weakly correlated; object_1 remains the main blocker and ACT alone is not yet stable full-score.
  - Seed2 finished 100k on 2026-05-22: `final.pt` SHA256 `36493deefaf05694351db776435c2d53d9fe4a274c4f95febfe0fe8b36fea652` scored `0.00` on seed11 (`logs/eval_task_e_seed2_final_seed11.log`); latest `best_loss.pt` SHA256 `f14c9d7c536c1d4156a8603064ec65cc593d8659147da175212fd8d7cbad2950` scored `9.00` on seed11 (`logs/eval_task_e_seed2_bestloss_96900_seed11.log`, only object_3 inside). Do not deploy seed2 over current XSA candidate.
  - Strategy judgment as of 2026-05-22: keep ACT/XSA as the best deployable baseline and finish seed2 to 100k for evidence, but do not assume more ACT iterations will reach stable 18. Prepare a pi0.5/OpenPI fine-tune branch using the existing HDF5 demos converted to LeRobot format; use GraspNet/SAM3 mainly as a grasp/data-generation aid unless object_1/2 contact stability is proven in ATEC Isaac with real observations/actions.
  - XSA final checkpoint had the best current evidence but is not full-score stable. Direct eval of `runs/act-task-e-rgb-xsa-100demos-30k-seed11/checkpoints/final.pt` scored `15/15/15` on seeds `11/12/13` (object_2 and object_3 inside, object_1 outside), with seed-11 video `logs/videos/task_e_act_eval/xsa_final_seed11_recheck.mp4`. The same checkpoint deployed as `demo/policy_act.pt` later scored `9` on seed 11 and `0` on seed 12, so there is residual nondeterminism/contact instability. Current `demo/policy_act.pt` SHA256 is XSA final `c3eacae3f1b0ec8ccde9fdc05d680fff119cc2ca70e1f79e4ddd5610c4bbfa93`; previous seed1 best was backed up as `demo/policy_act.seed1_best_backup_20260522_012719.pt`.
- pi0.5 / OpenPI Task E branch, 2026-05-22:
  - OpenPI repo/env: `/home/ubuntu/src/openpi-ebench-clean`, Python env `/home/ubuntu/envs/openpi-pi05`; run with `PYTHONPATH=/home/ubuntu/src/openpi-ebench-clean/src:/home/ubuntu/src/openpi-ebench-clean/packages/openpi-client/src`, `HF_LEROBOT_HOME=/home/ubuntu/projects/robotics_shared/datasets/lerobot`, `HF_HOME=/home/ubuntu/projects/robotics_shared/hf_cache`.
  - Avoid: do not use `uv run` in the OpenPI repo; it tries to create a large local `.venv` and download duplicated packages.
  - Converter: `scripts/pi05/convert_task_e_hdf5_to_lerobot.py` maps ATEC 8D Piper qpos/action to OpenPI fixed-base 16D padded `state/action.{joints,gripper}` and writes 3 identical 224x224 camera streams. 224 smoke dataset `atec/task_e_obj321_servo_smoke2_stride10_224` wrote 343 frames in about 70s and occupies 48M.
  - OpenPI configs added in `/home/ubuntu/src/openpi-ebench-clean/src/openpi/training/config.py`: `pi05_atec_task_e_smoke_224` and `pi05_atec_task_e_s5_224`. The formal `s5_224` config uses repo `atec/task_e_obj321_servo_100demos_s5_224`, batch size 1, 30k steps, save interval 30k to avoid 42G checkpoint spam.
  - Important fix: ATEC HDF5 `actions` are env actions `(joint_target-default)/0.5`, not absolute joint targets, so all ATEC OpenPI configs must keep `extra_delta_transform=False`. Using the OpenPI delta transform here would subtract state again and train on the wrong target.
  - Smoke verification passed: `pi05_atec_task_e_smoke_224` norm stats wrote under `/home/ubuntu/projects/robotics_shared/checkpoints/openpi/atec_runs/assets/pi05_atec_task_e_smoke_224/...`; 10-step train saved successfully once, but the temporary 42G checkpoint was deleted to preserve disk space. Log: `logs/pi05_atec_task_e_smoke_224_10step_20260522.log`.
  - Inference bridge added: `demo/solution_pi05.py` talks to an OpenPI websocket server and maps ATEC observation/action to/from the 16D fixed-base format. Eval entry point: `scripts/pi05/eval_task_e_pi05.py`. Default `ATEC_PI05_ACTION_REPEAT=5` matches the current stride-5/fps-10 training dataset. After a checkpoint exists, use `scripts/pi05/run_pi05_eval_checkpoint.sh <checkpoint_step_dir>` to start `scripts/serve_policy.py policy:checkpoint ...`, run seeds 11/12/13, and record a seed-11 video.
  - Eval summary helper: `scripts/pi05/summarize_pi05_eval.py <logs/pi05_eval_*>` parses seed logs for score/object basket status and is called automatically at the end of `run_pi05_eval_checkpoint.sh`.
  - Formal s5 conversion service started: unit from `logs/pi05_convert_task_e_s5_224.latest_unit`, log from `logs/pi05_convert_task_e_s5_224.latest`; auto continuation service `logs/pi05_s5_train_after_convert.latest_unit` waits for conversion, then computes norm stats and trains `pi05_atec_task_e_s5_224`. Eval watcher `logs/pi05_100demos_eval_after_train.latest_unit` waits for training, then calls `CONFIG=pi05_atec_task_e_s5_224 PORT=8017 scripts/pi05/run_pi05_eval_checkpoint.sh <latest_ckpt>`.
  - Early-signal branch: `pi05_atec_task_e_20demos_s5_224` config uses repo `atec/task_e_obj321_servo_20demos_s5_224`, 2k steps, batch size 1, no delta transform. Conversion service/log are in `logs/pi05_convert_task_e_s5_224_20demos.latest_unit` and `.latest`; auto continuation service `logs/pi05_20demos_train_after_convert.latest_unit` waits, computes norm stats, then trains. Eval watcher `logs/pi05_20demos_eval_after_train.latest_unit` waits for training, then calls `CONFIG=pi05_atec_task_e_20demos_s5_224 PORT=8016 scripts/pi05/run_pi05_eval_checkpoint.sh <latest_ckpt>`. Use this for quick pi0.5 viability checks before the 100-demo run finishes.
  - Note: the first 20-demo training run started from the old script name `atec_task_e_pi05_20demos_s5_224_10k_20260522070624`, but the loaded config was already patched to `num_train_steps=2000`; the name is cosmetic. The script was patched afterward to use `_2k_` in future run names.
  - 20-demo result: checkpoint `.../pi05_atec_task_e_20demos_s5_224/atec_task_e_pi05_20demos_s5_224_10k_20260522070624/1999` scored `0/0/0` on seeds 11/12/13. Logs: `logs/pi05_eval_20260522074137`; video: `logs/videos/task_e_pi05_eval/pi05_pi05_atec_task_e_20demos_s5_224_seed11_20260522074137.mp4`. The 42G 20-demo checkpoint directory was deleted after evaluation to preserve disk; dataset and logs were kept.
  - 100-demo formal run status, 2026-05-22 08:56 CST: conversion and norm stats are complete; training unit `atec-pi05-s5-train-after-convert-20260522065716.service` is actively running `scripts/train.py pi05_atec_task_e_s5_224 --exp-name atec_task_e_pi05_s5_224_30k_20260522065716 --overwrite`. Latest observed progress was about `731/30000` steps at roughly `3.3-3.6 it/s`, GPU util `100%`, ~80 GB VRAM. Log: `logs/pi05_train_task_e_s5_224_20260522065716.log`.
  - 100-demo checkpoint/eval gate: config `pi05_atec_task_e_s5_224` saves only at `30000` steps (`save_interval=30000`), expected final dir `.../checkpoints/pi05_atec_task_e_s5_224/atec_task_e_pi05_s5_224_30k_20260522065716/30000`. Eval watcher unit `atec-pi05-100demos-eval-after-train-20260522070900.service` waits for the training unit, then runs `CONFIG=pi05_atec_task_e_s5_224 PORT=8017 scripts/pi05/run_pi05_eval_checkpoint.sh <checkpoint>` for seeds 11/12/13 and records a seed-11 video.
  - Deployment decision gate: current deployed ACT/XSA baseline is still `demo/policy_act.pt` SHA256 `c3eacae3f1b0ec8ccde9fdc05d680fff119cc2ca70e1f79e4ddd5610c4bbfa93`. Do not replace it with pi0.5 unless the 100-demo pi0.5 checkpoint has real ATEC Isaac eval evidence on seeds 11/12/13 and is clearly better. Required completion artifacts for this branch are: final checkpoint dir, `logs/pi05_eval_*` with seed11/12/13 `[RESULT]` and `[BASKET]`, and `logs/videos/task_e_pi05_eval/pi05_pi05_atec_task_e_s5_224_seed11_*.mp4`.
  - Post-eval comparison helper: `scripts/pi05/compare_pi05_vs_act_baseline.py <logs/pi05_eval_*> --repo <repo>` compares pi0.5 seed logs with current ACT/XSA evidence (`eval_task_e_xsa_final_seed11/12/13_recheck.log` plus deployed seed11 recheck) and prints `HOLD_ACT_XSA` unless pi0.5 is clearly better. `scripts/pi05/run_pi05_eval_checkpoint.sh` calls this automatically after `summarize_pi05_eval.py`. It was sanity-checked on the 20-demo failed eval and correctly reported `HOLD_ACT_XSA`.
  - 100-demo result, 2026-05-22 11:27 CST: checkpoint `.../pi05_atec_task_e_s5_224/atec_task_e_pi05_s5_224_30k_20260522065716/29999` (about 42G) evaluated at `logs/pi05_eval_20260522112701` and scored seed11 `0.00`, seed12 `0.00`, seed13 `6.00` (mean `2.00`). Seed11 video: `logs/videos/task_e_pi05_eval/pi05_pi05_atec_task_e_s5_224_seed11_20260522112701.mp4`. The comparison helper printed `HOLD_ACT_XSA`; do not deploy pi0.5 over current XSA.
  - Monitoring: the old heartbeat automation `atec-task-e-seed2-resume-monitor` was deleted after the 100-demo pi0.5 train/eval completed, because the monitored branch is done. Recreate a new monitor only if starting another long run.
- Known-good environment:
  - Python: `/home/ubuntu/envs/genmanip-isaac5-py311/bin/python`.
  - Required env prefix: `OMNI_KIT_ACCEPT_EULA=YES PYTHONUNBUFFERED=1`.
  - Required `PYTHONPATH`: IsaacLab source dirs under `/home/ubuntu/envs/genmanip-isaac5-py311/lib/python3.11/site-packages/isaaclab/source/*`, plus `source/atec_rl_lab` and `scripts/act` from this repo.
- Current ACT baseline:
  - Object 3 only works with dataset `datasets/atec_task_e`, run `runs/act-task-e-rgb-100demos-seed1`, observed eval score about 6.
  - Baseline eval video: `logs/videos/task_e_act_eval/current_single_object_6pt.mp4`.
- Scripted oracle status:
  - Official/local Task E collector comment said object 3 works and object 1/2 need distance modifications.
  - Piper gripper joint limits observed on 2026-05-20: `joint7` is `[0.0, 0.035]`, `joint8` is `[-0.035, 0.0]`. Use legal close target `[0.0, 0.0]`, not `[-0.015, 0.015]`.
  - Object 2 had a previously observed basket candidate: offset `(dx=0.006, dy=0.046)`, `z=0.045`, `carry_z=TABLE_TOP_Z+0.075`, `place_z=TABLE_TOP_Z+0.075`, `target_y=-0.24`, `yaw=-pi/2`, keep gripper closed in transport. Treat this as not yet robust and likely push/slide rather than true lift: on 2026-05-20 a trace rerun showed the earlier search used candidate-index-dependent RNG, so success could be placement-specific.
  - On 2026-05-20 13:47, object 2 candidate search first 20 all failed under fixed trial seed. Best was cand 011 (`dx=0.030, dy=0.026, z=0.045, carry_z=TABLE_TOP_Z+0.075, target_y=-0.24, yaw=+pi/2`), final `(1.224,-0.009,1.213)`, `reward_lifted=True`, outside by `dy=+0.291`. Object 2 is not ready for demo collection/training.
  - Object 1 is unresolved as of 2026-05-20. Prior search candidates mostly pushed/slid the object; final basket failure alone was too weak to prove a real grasp. Add/keep per-stage diagnostics before trusting any object 1 candidate.
- Important local changes:
  - `scripts/act/eval_task_e_act.py` evaluates ACT and can record video.
  - `demo/solution_act.py` supports `ATEC_ACT_POLICY_PATH` and `reset_episode()`.
  - `scripts/act/cli_args.py` adds `--max_attempts`; `scripts/act/collect_demos_task_e.py` uses it and prints object basket status.
  - `scripts/act/task_e/collector.py` uses `_BASKET_MAX_Z = TABLE_TOP_Z + 0.15`, has `basket_status_lines`, and can return numeric trace diagnostics when `trace=True` including object z-gain, `gripper_base` distance, and `link7/link8` finger-body center/gap metrics.
  - `scripts/act/search_task_e_grasps.py` is the per-object candidate bank/search entry point.
  - `scripts/act/search_task_e_grasps.py` should keep trial seeds independent of candidate index so candidates are compared on the same object placements.
  - `scripts/act/search_task_e_grasps.py` prints `[BEST]` even when no candidate succeeds, using basket distance plus z-gain as a debugging score.
  - `scripts/act/search_task_e_grasps.py` supports `--start_candidate` to resume long searches without re-testing earlier candidates.
  - `scripts/act/search_task_e_grasps.py` now supports per-candidate `push_x/push_y/push_z` plus `close/lift/transport/place/open` step overrides. Use this to test trajectory timing without editing global config for each trial.
  - `scripts/act/task_e/state_machine.py` now has object-servo transport/place/open support for object 1/2: `OBJ_SERVO_TO_BASKET_STATES`, `OBJ_SERVO_XY_GAINS`, and `OBJ_SERVO_MAX_XY` in `scripts/act/task_e/config.py`. The TRANSPORT correction is ramped by state progress to avoid yanking objects out of the gripper.
  - `scripts/act/collect_demos_task_e.py` now disables default cameras and `cfg.observations.image` when `need_camera=False`; this fixes headless collection without `--enable_cameras`.
- Latest object 1 trace:
  - On 2026-05-20 13:19, foreground search candidates 1-12 all failed: `lifted=False`, `reward_lifted=False`, `z_gain` about `0.000-0.003m`; best candidate was 004 but still final `y=0.210`, far outside basket. Next search should resume at candidate 13 to test push/sweep fallback candidates.
  - On 2026-05-20 13:24, foreground search candidates 13-24 tested push/sweep fallback. No basket success yet, but contact is real: candidate 019 (`dx=0.040, dy=0.140, z=-0.020, carry_z=TABLE_TOP_Z+0.035, target_y=-0.240, yaw=0, push_y=-0.18`) reached `lifted=True`, `reward_lifted=True`, `z_gain=0.388`, final pos about `(1.276, -0.072, 1.260)`. It overshoots upward/outside basket, so next search should refine around push_y `-0.14..-0.18`, lower carry/place behavior, and possibly static place release.
  - On 2026-05-20 13:36, focused candidates 25-34 around candidate 019 all failed. Best was cand 034: `dx=0.040, dy=0.140, z=-0.020, carry_z=TABLE_TOP_Z+0.035, target_x=-0.17, target_y=0.0, push_y=-0.18`, final `(1.280,-0.064,1.257)`, `reward_lifted=True`, still outside by `dy=+0.236`.
  - On 2026-05-20 13:42, timing candidates with longer/lower `TRANSPORT` and `PLACE/OPEN` overrides all failed. Best was cand 002 with `transport_steps=900`, final `(1.280,-0.042,1.257)`, outside by `dy=+0.258`. This proves simply extending the push/sweep duration does not solve object 1; the contact geometry slips/throws the box before basket placement.
  - On 2026-05-20 14:15, object-servo ramp still failed object 1 first 8 candidates. Best was cand 003 (`push_y=-0.14`, `transport_steps=1200`), final `(1.294,-0.020,1.270)`, `reward_lifted=True`, outside by `dx=+0.214, dy=+0.280`. Object 1 still needs a different contact/grasp primitive, not more global servo.
- Latest object 2 trace:
  - On 2026-05-20 14:14, object-servo ramp made fixed-seed search candidate 001 succeed for object 2: final `(1.050,-0.272,0.865)`, `inside=True`, `z_gain=0.102`, candidate `{dx=0.006, dy=0.046, z=0.045, carry_z=TABLE_TOP_Z+0.075, target_y=-0.24, yaw=-pi/2}`.
  - On 2026-05-20 14:20, random `only_success` collection for object 2 still failed 0/4 attempts: final y errors remained about `+0.206..+0.310`, one early termination. So object 2 is promising but not robust enough for dataset/training yet.
  - On 2026-05-20 14:26, adding PLACE hold-until-near-basket still failed random object 2 collection 0/4: final y errors stayed about `+0.206..+0.334`.
  - On 2026-05-20 14:28, stronger object 2 servo (`gain=1.15`, `max_xy=0.36`, `TRANSPORT=480`) also failed 0/4 and worsened x drift; reverted object 2 to the softer fixed-seed-success servo values (`gain=0.85`, `max_xy=0.24`, no transport override). Next object 2 work should focus on grasp/contact robustness before transport, not stronger basket servo.
  - On 2026-05-20 14:41, object 2 multi-seed candidate validation (`--trials 3 --max_candidates 12`) found 0/12 robust candidates. The previous fixed-seed-success candidate failed immediately under this stricter check (`result=0/3`, final y error about `+0.262`). Best was yaw `0.0` with `z_gain=0.355` but final `(1.197,-0.001,1.277)`, outside by `dy=+0.299`. Treat object 2 as contact/grasp unstable, not transport-servo-limited.
- Current debugging judgment:
  - Do not integrate GraspNet first. For this small three-object simulated task, a reproducible scripted candidate bank with explicit lift/grasp diagnostics is faster and easier to pass code-review reproduction.
  - If object 1 remains blocked, collect/train object 2+3 first for a practical score bump, then return to object 1. Next object 1 work should change the motion primitive/contact mode, not merely sweep more static offsets or longer transport times.
- Useful verification commands:
  - Single object 1 debug video: `python scripts/act/collect_demos_task_e.py --pick_objects 1 --num_demos 1 --max_attempts 3 --headless --enable_cameras --only_success --save_video --video_dir logs/videos/task_e_oracle_obj1_fixed --output_dir logs/debug_task_e_obj1_fixed`.
  - Object 2+3 sanity check: `python scripts/act/collect_demos_task_e.py --pick_objects 2 3 --num_demos 1 --max_attempts 3 --headless --enable_cameras --only_success --save_video --video_dir logs/videos/task_e_oracle_obj23_fixed --output_dir logs/debug_task_e_obj23_fixed`.
- External reference checked:
  - `https://github.com/fiveages-sim/lerobot_ros2` is a public ROS2/LeRobot integration repo. It exposes generic ROS2 robot/camera wrappers and examples with `single_arm.pick/place` YAML skills, but it does not directly provide an ATEC Piper-compatible grasp planner. Its `robot_action_composer` submodule is private/inaccessible by HTTPS clone in this environment.

- Storage cleanup on 2026-07-19:
  - The final filtered three-object dataset and its 337 split files are preserved in `Datawhale/atec2026-task-e-reproducibility` on Hugging Face. The local `datasets/atec_task_e_obj321_servo_100demos/trajectory_filtered.hdf5` was deleted after remote verification.
  - The original 122G `trajectory.hdf5` was deleted after the collection and filtering scripts were uploaded under `code/task_e_collection/` in the same Hugging Face dataset repository. The raw trajectories themselves are not recoverable from those scripts.
  - The obsolete 30G `datasets/atec_task_e/` dataset, 15G `runs/` training intermediates, generated `logs/`, `logs/videos/`, and `submissions/` packages were deleted. Final model, reproducibility dataset, logs archive, source code, robot model, third-party code, and this memory remain available.
  - Local ATEC project size is now about 7.4G. Do not delete `third_party/anygrasp_sdk` or the source/model directories unless a separate cleanup decision is made.