ACT policy for PushT (200K steps)
Trained with LeRobot on lerobot/pusht
(human teleoperation demonstrations, 206 episodes).
Training
| Policy | ACT (CVAE + transformer, resnet18 backbone) |
| Steps | 200,000 (batch 64, 10 fps data) |
| Learning rate | 1e-05 |
| AMP | true |
| Observation | observation.image (96x96x3) + observation.state (2) |
| Action | 2D pusher target position, chunk_size 100 |
Evaluation (2026-09-09)
Protocol identical across checkpoints: gym_pusht/PushT-v0, pixels_agent_pos, 300 steps max, 200 episodes, seeds 1000-1199, batch 50, async envs.
| Checkpoint | Success rate | avg max reward (coverage) | avg sum reward |
|---|---|---|---|
| 20K steps | 0/200 (0.0%) | 0.367 | 34.0 |
| 200K steps (this model) | 4/200 (2.0%) | 0.419 | 38.8 |
Wilson 95% CI for success rate: [0.8%, 5.0%].
Full per-episode results, report and 200 rollout videos: zyh1212zyh/act_pusht_200k (dataset) · Training curves: https://wandb.ai/7bread-lab/lerobot
Usage
lerobot-eval --policy.path=zyh1212zyh/act_pusht_200k --env.type=pusht --eval.n_episodes=50
- Downloads last month
- 8