DP_sim_click_bell_ft8000-lr1e4 (checkpoint 008000)

Diffusion Policy for the RoboSynChallenge simulation task click_bell. The repository root is a complete LeRobot pretrained_model directory.

Files

File	Size (bytes)	sha256
config.json	2472	808671a8249a002a9c77c1b8293b535ea7a2243b200d086547430bf3e3754a4b
model.safetensors	1083151064	a74fe4b2dcd23db11c99769f7de87b81e650f74955aa812dfb5c873e8dda37a9
train_config.json	8548	c0bdfd17f107ff45500373b6d78297e85521733836145cfa4595019a3f0e15e4
finetune_provenance.json	3192	678df45d78a672b816aafe225c7d732d215ede32a0f597adc5da75183774fc92

Model

• Architecture: Diffusion Policy (DDPM, epsilon prediction, 100 train timesteps; 1D U-Net down_dims [512, 1024, 2048], ResNet18 visual backbone)
• Initialization: official released checkpoint RoboSynChallenge/DP_sim_click_bell @ fb8c9551e0fade7c4888a926ab577b85f0684da1
• Training: lr 1e-4 (Adam), batch size 64, 10000 steps configured (delivered checkpoint at step 8000), seed 1000; normalization: preserve official checkpoint tensors
• Data: cobotmagic_Sim_click_bell, LeRobot v2.1, 1000 episodes, 74000 frames, 25 fps. The data is not included.
• W&B run: run id txq5iahe (project cyanrobosyn/robosyn-overnight, offline mode)
• Code: robosyn5090 team clone; experiment-local fine-tune tooling (finetune_official.py, train_job.py, run_finetune.sh)
• Environment: Python 3.10.12, lerobot 0.3.3, torch 2.7.0+cu128, RTX 5090

Input and output

• State inputs: observation.state, observation.qvel, observation.qf [14], n_obs_steps 2
• Image inputs: observation.images.cam_high, observation.images.cam_left_wrist, observation.images.cam_right_wrist [3, 480, 640]
• Action: action [14] absolute joint positions (CobotMagic)
• Normalization: VISUAL MEAN_STD, STATE MIN_MAX, ACTION MIN_MAX
• horizon 32, n_action_steps 32

Procedure: download and evaluate

1. hf download mair-robosyn/DP_sim_click_bell_ft8000-lr1e4 --repo-type model --revision 245a2377ed113f4d45d1562ed626f95b7f6c0134 --local-dir checkpoints/DP_sim_click_bell_ft8000-lr1e4 (model files verified at this revision; later commits update this model card only)
2. bash policy/dp/eval.sh click_bell random checkpoints/DP_sim_click_bell_ft8000-lr1e4 0 --pytorch_device cuda --headless True

Results (team-internal, not organizer results)

• Protocol: setting random, seed 240926, host robosyn5090
• Success rate: 20 episodes: 15/20 = 75.0%; 50 episodes: 37/50 = 74.0% (mean action steps 132.5)
• Reference: checkpoint 010000 of the same run, 20 episodes, same protocol: 14/20 = 70.0%

Status

• UPLOADED: yes, revision 245a2377ed113f4d45d1562ed626f95b7f6c0134 (re-download sha256 match for all 4 files, 2026-10-10)
• LOAD VERIFIED: yes, same revision (2026-10-10). Minimal load + dummy batch: action (1, 14) torch.float32 finite True.
• EVAL VERIFIED (team-internal): yes, same revision, 5 episodes (seed 240926, max_steps 500), SR 3/5 = 60%, 2026-10-10. Official evaluation bridge (official eval loop, deferred simulator close) on the re-downloaded directory wrote evaluation_metrics.json.

Warning

These results come from our local runs. The organizers did not verify them.