ECM27 Sim Track ACT checkpoints
Three ACT training checkpoints based on the corrected H5 demonstrations in rocochallenge2025/ecm2027_sim_data. The earlier aggregate LeRobot export was excluded.
Training setup
- Dataset conversion: 173 successful, non-empty episodes; 155 train / 18 validation episodes; 894,635 frames total.
- Rate: 30 FPS.
- Inputs: three 224x224 RGB camera views and a 77-dimensional state.
- Output: 20-dimensional action.
- Policy: LeRobot ACT, chunk size 100, one action step.
- Training: 50,000 steps, batch size 16, learning rate 1e-5, seed 1000.
- Stack: LeRobot 0.6.1, PyTorch 2.11, CUDA 12.8.
- Video timestamp tolerance: 0.001 seconds.
Checkpoints and validation loss
| Checkpoint | Validation loss |
|---|---|
| checkpoint-40000 | 0.3723 |
| checkpoint-45000 | 0.3738 |
| checkpoint-50000 | 0.3761 |
The lowest recorded validation loss was 0.3697 at step 49,000; checkpoint saves were every 5,000 steps, so that exact step is not included. Final training loss was 0.222. Validation loss is the held-out dataset objective, not a task success rate.
Each checkpoint directory includes pretrained_model/ for policy inference and training_state/ for resuming training.