BASELINE: LeRobot stock ACT on the bussing task
⚠️ This is a reference baseline, NOT the INSAIT whiteboard policy. It is LeRobot's off-the-shelf ACT implementation, included only as supporting evidence for comparison. The whiteboard policy is a different architecture: frozen swappable ViT encoder (DINOv2 / SigLIP / JEPA, 86.6M), 7-layer decoder, optional language conditioning, 142.3M parameters — versus this model's trainable ResNet-18, 1-layer decoder, no language, 51.6M.
LeRobot's stock ACT policy trained on Sombit/bussing_task_new. Loads with
ACTPolicy.from_pretrained() — nothing custom to install.
from lerobot.common.policies.act.modeling_act import ACTPolicy
policy = ACTPolicy.from_pretrained("oku006/bussing-act")
Interface
observation.images.main |
320×180 RGB, overhead camera |
observation.images.wrist |
320×180 RGB, wrist camera |
observation.state |
8-d: 7 joint angles + gripper |
| action | 8-d: 7 joint velocities (rad/s) + gripper position |
| chunk_size | 16, n_action_steps 8 |
Same action space as the π0.5 checkpoint (oku006/bussing-pi05-droid), so both
drop into the same runner with action_space: joint_velocities.
Set flip_gripper: False — trained on the raw recorded convention
(1.0 = open, ~0.3 = closed on an object).
Measured, held-out episodes [2, 4, 10, 20, 21]
| metric | value |
|---|---|
| velocity error | 0.0203 rad/s |
| zero-motion baseline | 0.0267 rad/s |
| beats doing nothing by | 24.1% |
| motion ratio (1.0 = matches human) | 0.940 |
| joint error over the 8 executed steps | 0.0307 rad (1.76°) |
| clipped by the controller | 2.1% |
Trained 40k steps (~50 epochs) on 33 of 38 episodes, loss 0.098. Checkpoints also exist at 10k/20k/30k steps.
Limits
33 episodes of one task. No generalization to new objects or layouts expected. All recordings start from ≈ the same home pose (1 cm spread), so starting far from it is out of distribution.
- Downloads last month
- -