{ "algorithm": "supermimic", "actor_input_normalization": "rsl_rl_empirical_real_next_obs", "task": "box_in_bin", "seed": 1, "smoke": false, "reset_bank": "/mnt/data3/ZhouZhiqiang/supermimic_manipulation/banks/box_in_bin.pt", "dataset": { "contract": { "task": "box_in_bin", "robot": "arx5", "joint_order": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "joint8" ], "frame": "robot_base_local", "pose": "xyz_wxyz", "joint_units": [ "rad", "rad", "rad", "rad", "rad", "rad", "m", "m" ], "joint_velocity_units": [ "rad/s", "rad/s", "rad/s", "rad/s", "rad/s", "rad/s", "m/s", "m/s" ], "position_unit": "m", "box_velocity": "local_linear_xyz_angular_xyz", "box_velocity_units": [ "m/s", "m/s", "m/s", "rad/s", "rad/s", "rad/s" ], "categories": [ "ObjectAnywhereEEAnywhere", "ObjectRestingEEGrasped", "ObjectAnywhereEEGrasped", "ObjectPartiallyAssembledEEGrasped" ] }, "provenance": { "workspace": { "frame": "world", "xy_min": [ 0.2, -0.2 ], "xy_max": [ 0.5, 0.3 ], "containment": "complete_object_geometry", "units": "m" }, "proposal_workers": 64, "proposal_seed_schedule": "base_seed_ordered_chunk_seedsequence", "proposal_chunking": "balanced_contiguous_chunks_up_to_worker_count", "method": "procedural_mjlab_contact_hold", "candidate_source": "mujoco_fk", "physics_randomization": false, "compiled_physics": { "gripper": { "force_units": "N per finger", "actuators": { "joint7_position": { "forcerange": [ -5.0, 5.0 ], "forcelimited": true, "ctrlrange": [ 0.002, 0.044 ], "ctrllimited": true, "gainprm": [ 200.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ], "biasprm": [ 0.0, -200.0, -50.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ], "gaintype": 0, "biastype": 1 }, "joint8_position": { "forcerange": [ -5.0, 5.0 ], "forcelimited": true, "ctrlrange": [ 0.002, 0.044 ], "ctrllimited": true, "gainprm": [ 200.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ], "biasprm": [ 0.0, -200.0, -50.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ], "gaintype": 0, "biastype": 1 } } }, "gripper_contact": { "friction_order": [ "sliding", "torsional", "rolling" ], "geoms": { "robot/link7_collision": { "body": "robot/link7", "friction": [ 2.0, 0.005, 0.0001 ], "priority": 0, "condim": 3, "contype": 1, "conaffinity": 1 }, "robot/link8_collision": { "body": "robot/link8", "friction": [ 2.0, 0.005, 0.0001 ], "priority": 0, "condim": 3, "contype": 1, "conaffinity": 1 } } }, "contact_solver": { "cone": "elliptic", "impratio": 5.0 }, "allocation_limits": { "nconmax": 24, "njmax": 128 }, "parameter_scope": "nominal compiled model; startup physics randomization disabled" }, "hold_command": "fixed_initial_ee_pose_via_relative_osc_action", "replay_validation": { "steps": 20, "required_stable_steps": 20, "counts": { "ObjectAnywhereEEAnywhere": { "tested": 19905, "passed": 19905 }, "ObjectRestingEEGrasped": { "tested": 19830, "passed": 19800 }, "ObjectAnywhereEEGrasped": { "tested": 19486, "passed": 19368 }, "ObjectPartiallyAssembledEEGrasped": { "tested": 19024, "passed": 18982 } }, "saved_state": "input_before_replay", "command": "fixed_initial_ee_pose_and_same_gripper", "seed_offset": 700000 }, "floating_grasp_source": "mujoco_fk", "seed": 42, "hold_steps": 10, "hold_seconds": 0.5, "accepted_consecutive_steps": 10, "stable_seconds": 0.5, "control_dt": 0.05, "physics_dt": 0.0020833333333333333, "decimation": 24, "controller": { "entity_name": "robot", "clip": null, "arm_joints": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6" ], "finger_joints": [ "joint7", "joint8" ], "ee_body": "link6", "closed_position": 0.002, "open_position": 0.044, "effort_limit": 50.0, "inertia_solver": "pinv", "inertia_rcond": 1e-10, "uncouple_pos_ori": true, "position_scale": 0.02, "orientation_scale": 0.2, "motion_stiffness": [ 200.0, 200.0, 200.0, 300.0, 300.0, 300.0 ], "motion_damping_ratio": [ 3.0, 3.0, 3.0, 1.0, 1.0, 1.0 ], "input_limit": 12.0, "linear_velocity_limit": 1.5, "angular_velocity_limit": 8.0 }, "num_envs": 10000, "max_proposals_per_batch": 100000000, "robot_table_penetration_limit_m": 0.001, "initial_box_robot_contact_filter": true, "ungrasped_box_xy_source": "collision_checked_resting_grasp" }, "split": "train", "counts": { "ObjectAnywhereEEAnywhere": 10000, "ObjectRestingEEGrasped": 10000, "ObjectAnywhereEEGrasped": 10000, "ObjectPartiallyAssembledEEGrasped": 10000 } }, "run": { "iterations": 10000, "buffer_steps": 1000, "warmup_steps": 64, "world_model_batches": 30, "world_model_only_iterations": 1000, "updates_per_sim_step": 0.4, "learning_rate_decay_iterations": 10000, "save_interval": 1000, "log_interval": 50, "action_noise": 0.1, "compile_world_model": true, "compile_policy_update": true, "pin_replay": false, "streaming_buffer_steps": 1000, "world_model_schedule": "epochs", "world_model_epochs": 4.0, "validation_fraction": 0.05, "validation_batches": 4, "validation_recent_steps": 100, "exploration_fraction": 0.0, "exploration_start_probability": 0.05, "exploration_steps": 10, "policy_update_interval": 1 }, "imagination_overrides": {}, "world_model_validation_envs": 205, "agent": { "wm_hidden": [ 2048, 2048, 2048, 2048, 2048 ], "policy_hidden": [ 2048, 1024, 512, 256 ], "critic_hidden": [ 256, 128 ], "wm_window": 8, "horizon": 16, "wm_batch_size": 2048, "policy_batch_size": 4096, "learning_rate": 0.0002, "gamma": 0.99, "lambda_return": 0.95, "critic_epochs": 5, "critic_minibatches": 4, "max_grad_norm": 1.0, "policy_grad_value_clip": 0.1, "normalizer_max_count": 200000000.0, "amp": "bf16", "tf32": true, "wm_ensemble_size": 1, "uncertainty_penalty": 0.0 }, "device": "cuda:0", "device_name": "NVIDIA A100-SXM4-80GB", "torch": "2.9.1+cu128", "cuda": "12.8", "precision": { "parameters_normalizers_state_reward": "float32", "mlp_autocast": "bf16", "controller_ik": "float64", "matmul_allow_tf32": true, "cudnn_allow_tf32": true }, "torch_compile": true, "torch_compile_scope": "world_model_rollout_loss", "torch_compile_options": { "fullgraph": true, "dynamic": false, "mode": "default" }, "torch_compile_policy_update": true, "logger": "tensorboard", "cuda_graph_mode": "plain", "timeout_bootstrap": false, "resume_is_exact": false, "resume_checkpoint": null, "evaluation": "separate complete-episode workflow (clab eval policy)", "limitations": [ "Bank controller match is not a physical hold/capacity certificate", "World model is a reduced stochastic surrogate, not an exact MuJoCo state", "Imagination reward/success is analytic from endpoint states, not a learned reward head", "Endpoint imagination reward/success is approximate and need not match native contacts", "Real collection and evaluation retain native rewards and terminal conditions" ], "reward_mode": "imagined_endpoint_geometry_velocity_release_shaped", "reward_matches_native": false, "cuda_graph_runtime": { "enabled": true, "graph_conditional": false, "step_graph_created": true, "forward_graph_created": true, "reset_graph_created": true } }