WDong/so101-act-erythromycin-tea-ablation

SO-101 ACT erythromycin-on-tea ablations

Five ACT policies trained for an SO-101 pick-and-place task:

Pick up the red erythromycin ointment box and place it on top of the green Rizhao tea tin.

The demonstrations contain two 640x480 camera streams (fixed, wrist) and six calibrated joint-position dimensions in this order:

1. shoulder_pan.pos
2. shoulder_lift.pos
3. elbow_flex.pos
4. wrist_flex.pos
5. wrist_roll.pos
6. gripper.pos

Results

Held-out episode IDs are 20, 47, and 89. The table reports mean absolute error after decoding predictions into calibrated absolute joint space. Lower is better.

ID	Training profile	Action representation	LR	Val-best step	K=5 MAE	K=10 MAE	K=5 shoulder-pan signed error
T1	both-trim clean112	absolute	warmup + cosine	32,500	3.777	4.234	-0.361
T2	front-trim clean112	absolute	warmup + cosine	27,500	3.621	4.021	-0.250
T3	both-trim full117	absolute	warmup + cosine	35,000	3.668	4.126	-0.141
T4	both-trim full117	relative arm5 + absolute gripper	warmup + cosine	37,500	2.872	3.418	-0.797
T5	both-trim clean112	absolute	constant 1e-5	57,500	3.692	4.051	-0.030

T4 has the best offline absolute-space MAE, but also the largest negative shoulder-pan bias. T5 is the low-bias absolute-action baseline. T3 tests whether retaining recovery demonstrations helps closed-loop correction. T2 retains terminal static frames, so its temporal evaluation window contains 483 frames rather than the 439 frames used by the both-trim policies; its MAE should not be treated as a perfectly apples-to-apples win over the other absolute policies.

These are offline behavior-cloning metrics on logged observations, not autonomous rollout success rates.

Repository layout

Each T*_* directory contains:

• pretrained_model/: the selected validation-best LeRobot ACT checkpoint.
• curves/smolvla_curves.png: training and validation curves (the historical filename is retained by the plotting pipeline).
• curves/curve_summary.json and exported metrics.
• unified_heldout_eval.json: offline K=5/K=10 diagnostics.
• launch_manifest.json: training provenance.

Deployment semantics

• T1, T2, T3, and T5 output calibrated absolute joint-position targets.
• T4 predicts the first five dimensions relative to the same current observation state; the gripper remains absolute. The deployment adapter must decode the first five dimensions back to absolute targets exactly once. Do not apply a cumulative sum.
• Always verify calibration, joint order, camera order/orientation, output shape, finite values, limits, and first-action deltas with motors disabled before commanding hardware.
• Start real-robot evaluation with a short executed prefix such as K=5 and continuous human supervision.

The recommended test order is T4, T5, T3, followed by T1/T2 as secondary ablations.