Bigenlight/act_banana_in_pot_ee

ACT · banana-in-pot · EEF (10-D) — checkpoint 40k

Action Chunking Transformer (ACT) trained on the end-effector (EEF) action space for the task "put the right banana in the pot" (UR7e arm, GELLO teleoperation, LeRobot v3.0). This is the 40k-step checkpoint, selected as best by open-loop MAE.

This is the EEF counterpart of the joint-space model Bigenlight/act_banana_in_pot.

Action / observation space

• observation.state / action: 10-D = [x, y, z, r1..r6 (Zhou 6D rotation), gripper] — absolute next-frame TCP pose (xyz in metres) + gripper. (The joint model uses 7-D [q1..q6, gripper].)
• Cameras: observation.images.cam1, observation.images.cam2 (RGB, resized 360×640).
• Backbone: ResNet18 + VAE, chunk_size=100, ~51.6M params. Normalization: MEAN_STD.

Training

• Recipe identical to the joint baseline train_act_valdiag.sh except dataset + steps: --dataset.eval_split=0.117 (held-out episodes 45–50), batch 8, seed 1000, 50k steps.
• Dataset: banana_in_pot_ee_action (51 eps / 21,524 frames, 30 fps), built from the raw Bigenlight/banana_in_pot_raw via recorded tcp_pose (no FK needed).
• Hardware: single RTX A4000, ~2h43m. No overfitting (held-out eval_loss monotone to 0.4594@50k).

Held-out results (open-loop, eps 45–50)

checkpoint	pose MAE (m + 6D)	gripper acc
40k (this)	0.05564	0.914
50k	0.05564	0.911

Selected by open-loop MAE (repo convention), not by eval_loss.

⚠️ Note: EEF pose MAE mixes metres (xyz) and unitless 6D-rotation and is not directly comparable to the joint model's radian MAE. See the comparison writeup.

Usage

from lerobot.policies.act.modeling_act import ACTPolicy
policy = ACTPolicy.from_pretrained("Bigenlight/act_banana_in_pot_ee")

Links

• Experiments repo, full report, reproducibility (Docker): GitHub Bigenlight/banana-in-pot-experiments — see docs/ACT_EE_RESULTS.md and docs/JOINT_VS_EEF_ACT_COMPARISON.md.

License: CC-BY-NC-4.0 (trained on real-lab teleoperation video).