Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
Paper • 2304.13705 • Published • 7
ACT — dg5f_ur5e_newcablev5 (no GelSight)
ACT policy trained with LeRobot v0.6.0 on Kaz55/dg5f_ur5e_newcablev5, with the tactile (GelSight) cameras and the velocity/effort observations removed.
Inputs / outputs
| key | shape | |
|---|---|---|
| input | observation.state |
(26,) — UR5e ×6 + DG-5F ×20 |
| input | observation.images.realsense1 |
(3, 480, 640) |
| input | observation.images.realsense2 |
(3, 480, 640) |
| output | action |
(26,) |
Dropped from the source dataset: observation.images.gelsight1, observation.images.gelsight2,
observation.velocity, observation.effort.
Training
- 100,000 steps, batch size 8, AdamW (LeRobot ACT training preset), single RTX 4090, ~1 h 56 min
chunk_size=60,n_action_steps=1,temporal_ensemble_coeff=0.01- Data: 89 episodes / 96,348 frames at 30 fps (≈8.3 epochs)
- Final training loss ≈ 0.13 (from 8.19 at step 200)
- 51.6M parameters
No held-out evaluation was run; the loss above is training loss only.
Usage
from lerobot.policies.act.modeling_act import ACTPolicy
policy = ACTPolicy.from_pretrained("Kaz55/act_dg5f_ur5e_newcablev5_nogelsight")
- Downloads last month
- -