ACT_sim_water_pouring

ACT policy for the RoboSynChallenge simulation task water_pouring. This repository root is a complete LeRobot pretrained_model directory.

Team repository: mair-robosyn/ACT_sim_water_pouring. The checkpoint has been uploaded. The weights are unchanged from the upstream release; only pretrained_backbone_weights is set to null for offline backbone loading.

Local preparation checks passed. Fresh-download loading and simulator evaluation of the team HF copy remain unverified.

Files

File	Size (bytes)	sha256
config.json	1668	10596bd971c6e547d44a94fb92b3809c07d176dc72969d3e36a0f9aad0009df8
model.safetensors	206563032	40aa483bf3ff18314e89c17ed91066cca867a8aa587c5fb9ecbb963c0eed330c
train_config.json	5490	a6af1a8551ac61c0bfefd5dd79b0d67b78c156e6e87d606720a4bca1073db237
SmolVLA pre/postprocessors	N/A (ACT)	N/A

README.md is the model card. The original source is retained separately. The weight and train_config files are byte-identical to the pinned upstream release. Only pretrained_backbone_weights in the staging config changes from ResNet18_Weights.IMAGENET1K_V1 to null, as required by the team SOP. No ResNet download file, data, observations, videos, optimizer state or credentials are in this directory.

Model

• Architecture: ACTPolicy, ResNet18, dim_model 512, 8 heads, feedforward 3200, 4 encoder layers, 1 decoder layer; VAE latent_dim 32, 4 VAE encoder layers, KL weight 10.
• Initialization/source: RoboSynChallenge/ACT_sim_water_pouring @ 0bf0fcfc931a69c52871385f28068fcf873cf07a. This is the exact released model linked to the organizer's 72/100 result. No team fine-tuning or correction data was applied to these weights.
• Training (inherited config): lr 1e-5, lr_backbone 1e-5, batch 32, configured steps 80000, seed 1000, AdamW weight_decay 1e-4, AMP false. The release does not independently identify the actual saved optimizer update count or a numeric checkpoint directory; configured steps are not proof of completed steps.
• Data: RoboSynChallenge/cobotmagic_Sim_water_pouring, inspected LeRobot v2.1 metadata: 1000 episodes, 207000 frames, 25 Hz, revision 932a287657983c41465d6c9baee5dd478ef61b90. The original train_config leaves dataset revision null; the inspected metadata revision is not asserted as the original training revision. Data is not included.
• W&B run: inherited project robosynchallenge, run_id e8kr6qum, entity null. Exact URL N/A (not recorded).
• Code: official https://github.com/EDEM-AI/RoboSynChallenge @ 8818c82df5fa2522809b8105e7018950dcc9a9a8 inspected for current loading/evaluation; team docs @ 19f0d0b04c65194c6a8652e967a083d3bd24e071. Original training code commit N/A (not recorded). No upstream policy or task source was changed for packaging.
• Environment: local CPU check on ssh lp, existing sim/ACT Python 3.10.22, LeRobot 0.3.3, PyTorch 2.7.0+cu128. CUDA was uninitialized; no new environment, driver change, training or simulator episode was used.

Input and output

• Dataset: observation.state, observation.qvel, observation.qf, and action float32 [14]. All three cameras are RGB [480,640,3], 25 Hz.
• Saved normalizer: all three declared state fields have mean/std [14]; each RGB mean/std [3,1,1]; action mean/std [14]. All 16 buffers are unchanged, aggregate SHA256 e2e33f3c05086fee29bf38dde504cbe3c0069ed548d06a2b5856d378502a22c1.
• Runtime deploy input: observation.state [1,14] from normalized robot/qpos, plus observation.images.cam_high, cam_left_wrist, cam_right_wrist, each [1,3,480,640], float32 RGB scaled to 0..1. The ACT effective robot state is observation.state; the declared qvel/qf features are not supplied by the ordinary deploy adapter. The real observation check passed with only these four fields. No bottle/cup pose or other privileged information is a policy input.
• Action: [1,14], absolute targets ordered left arm 6 joints + left gripper + right arm 6 joints + right gripper. Arms use radians; gripper command positions use the original task action interface (meters). Commands are passed unchanged to env.step; no correction-controller mapping, takeover, or new binarization is used.
• Normalization: VISUAL, STATE, ACTION = MEAN_STD; saved checkpoint stats retained.
• chunk_size 50, n_action_steps 50; temporal ensemble null. Ordinary adapter act_step 10 means 10 environment actions per adapter call; it does not reduce the internal 50-action queue.

Procedure: download and evaluate

Before upload, the prepared directory can be checked on lp without a GPU:

CUDA_VISIBLE_DEVICES='' HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 \
/home/yizhenyu/robosyn-simdagger-a100-pilot/env/sim/bin/python \
/home/yizhenyu/water-official72-delivery-20261011/strict_check.py \
/home/yizhenyu/water-official72-delivery-20261011/staging/ACT_sim_water_pouring_official

After manual upload, record the actual 40-character HF commit. On a host with an existing compatible ACT/simulator environment and a clean official checkout at 8818c82df5fa2522809b8105e7018950dcc9a9a8:

hf download mair-robosyn/ACT_sim_water_pouring --repo-type model \
  --revision 1003b2e793cf38ffeb0735419e55bedfe8985a59 \
  --local-dir checkpoints/ACT_sim_water_pouring

bash policy/act/eval.sh water_pouring random \
  checkpoints/ACT_sim_water_pouring 0 \
  --max_episodes 5 --max_steps 500 --seed 240926 \
  --pytorch_device cuda --headless True \
  --eval_result_dir eval_result/water_hf_smoke

WATER_HF_REVISION must be the commit from the upload receipt, not the upstream commit. The evaluation command is the SOP five-episode pipeline check, not an exact reproduction of the organizer's historical 72/100 run. Check resource availability first. It has NOT been executed in this delivery. The running ACT003000 evaluation is preserved.

Official published reference (not our reproduction)

• Primary source: official results JSON and its README.
• water_pouring, ACT, pinned upstream 0bf0fcfc931a69c52871385f28068fcf873cf07a: 72/100 = 72%, simulation random, task config configs/water_pouring/random/gym_config.json, max 500 actions, average action steps 276.00.
• The source is schema 4, published 2026-08-19, recalculated 2026-08-21. It explicitly retains historical ACT success counts and task-limit-extrapolated step metrics. The document-level protocol revision is 755af40a00492f43bd4b9ead68b0def456b7c8b1; the README specifically pins the fresh DP evaluation to it. It does not establish the original ACT rollout code commit, exact seed list, feasibility filtering, or rollout hardware. Those details remain unverified.
• The RTX 5090 synthetic-observation ACT timing benchmark is separate from the historical success rollouts. Do not label that timing platform as verified hardware for the 72/100 success experiment. DP water is 36/100, not this 72/100 result.

Results (team-internal, not organizer results)

• Prepared official copy: strict CPU load passed, all saved tensors equal; finite dummy action [1,14], chunk [1,50,14]; an existing real observation with state + three cameras also gave finite [1,14]. Zero optimizer updates, CUDA uninitialized, zero new simulator episodes. This is a local preparation check, not the SOP post-upload LOAD VERIFIED status and not a task success-rate test.
• Historical official-source local record: 19/30 = 63.33%, random, 500 actions, seed generator 240926, 30 accepted of 72 expert-feasibility candidates; mean 302.0667 actions. Evidence water-recovery-20261011/original30_metrics.json and its source weight hash audit. It is a filtered historical source-checkpoint run, not a fresh HF mirror validation and not a reproduction of 72/100. Accepted seeds: 290298465, 942302943, 1165167967, 1751217874, 1364265482, 917421052, 209888944, 783217757, 162260320, 1853118987, 790779347, 450785757, 1624254241, 436607482, 630153513, 601867212, 188790587, 283940036, 537378871, 1584445822, 2072979752, 1521940945, 113606605, 1433950075, 346068329, 1691465875, 227740116, 2022641686, 969253029, 129467814.
• Separate ongoing ACT003000 experiment snapshot at 2026-10-11T20:59:45.418553+08:00: 26/51 = 50.98% on normally completed new registered seeds; 25 normal task failures, 1 terminal runtime error. It uses lp A30, unfiltered random seeds, 500 actions, pure ACT, no expert takeover. It is incomplete; final_report.json was absent at the snapshot. Its weight is f6ce67e0ca415aa002db6fcf62b29d8048a37f1ff14673cbf96253dd93cc2ffc, not this official weight. Its errors are excluded from the normal-completion SR denominator.

Status

• Package prepared: yes; local staging CPU check passed.
• UPLOADED: yes; team checkpoint files are available at mair-robosyn/ACT_sim_water_pouring, revision 1003b2e793cf38ffeb0735419e55bedfe8985a59. Fresh-download hash verification remains pending.
• LOAD VERIFIED: no in the SOP's fresh pinned team-HF-download sense; local CPU strict preparation check passed.
• EVAL VERIFIED (team-internal): no; no official five-episode run on a fresh team-HF download.
• SUBMITTED (website): no. No complete current ten-task submission manifest was found in the inspected team docs, local delivery records, or HF listing. No other task selection was changed. The website is https://robosyn-bench.net/; no form was sent.

Warning

The 72/100 figure is the organizer's published historical reference, not a team reproduction. The local 19/30 record and ongoing ACT003000 result use different protocols/seeds. This delivery does not prove superiority over ACT003000 or constitute a SimDAgger experiment. lp direct HF access failed certificate verification; no login, token output, TLS bypass or system-network change was performed. Manual upload, immutable team revision, public redownload and SOP load/eval checks remain pending.