Suyang99/xlerobot-act-cup-grasp-right

ACT — right-arm cup grasp (XLeRobot)

Trained with LeRobot ACT on Suyang99/xlerobot-cup-grasp-20260820-0230.

Task	Pick up the cup and place it down — right arm only (arms: 1.0)
Dataset	50 episodes / 19,453 frames / 30 fps, 5 cup positions
Policy	ACT, 51.6 M parameters
Checkpoint on main	step 20 000 of 20 000 (training ran to completion)
Earlier revision	step 4 000, kept at revision a920f833
Trained on	Jetson Orin Nano Super (8 GB unified), torch 2.8.0, batch size 2
Cameras	head, left_arm_wrist, right_arm_wrist
Action / state	17-dim (left arm 6 · right arm 6 · head 2 · base 3)

A previous version of this card read step 200001211111. That was an upload-script bug (tr -dc '0-9' concatenated every digit in training_step.json), not a real step count. Fixed.

Training did not converge — it was cut off while still improving

Read this before comparing the checkpoint against anything.

step	loss	l1_loss	kld	grad norm	epoch
100	11.630	0.413	1.1220	378.0	0.01
2 000	1.926	0.249	0.1680	110.0	0.21
6 000	0.664	0.168	0.0500	57.2	0.62
10 000	0.310	0.137	0.0170	34.1	1.03
14 000	0.196	0.115	0.0080	23.4	1.44
18 000	0.159	0.107	0.0050	18.8	1.85
20 000	0.138	0.094	0.0040	17.3	2.06

The curve is healthy — monotone, no divergence, no oscillation, gradient norm down 95%. But over the last 5 000 steps the mean loss still fell 13.2% (0.1750 → 0.1519). A converged run is flat there. This checkpoint stopped where it was still learning.

The reason is the step budget: batch_size = 2 × 20 000 steps = 40 000 samples over a 19 453-frame dataset — 2.06 epochs total. ACT is normally trained far longer.

No validation was run

eval_freq was None, so every number above is training loss. Nothing here measures overfitting or generalisation, and loss: 0.138 must not be read as "the policy learned the task". The project's own guidance is explicit that training loss is not an acceptance criterion.

The generalisation evidence that does exist comes from the sibling SmolVLA run (Suyang99/xlerobot-smolvla-cup-grasp-right), which was given a held-out cup position: unseen-position loss 0.2300 vs trained-position 0.2348 (0.98×). That is offline replay of recorded observations, not a closed-loop rollout.

KLD collapsed

kl_weight = 10.0, yet KLD fell from 1.122 to 0.004. The CVAE latent is effectively unused and the model has degenerated to deterministic chunk regression. Not unusual for ACT on a small single-task dataset, but worth knowing before treating the latent as meaningful.

Checkpoint availability

Ten checkpoints were written during training (every 2 000 steps). Eight no longer exist — they were deleted from the Jetson under disk pressure and never uploaded. Only these two survive:

step	where
4 000	revision a920f833 of this repo
20 000	main

hf download Suyang99/xlerobot-act-cup-grasp-right \
  --revision a920f833 --local-dir ./act_step4000

(huggingface-cli is deprecated and on recent versions prints help and exits 0 without downloading anything — use hf, or snapshot_download(..., revision="a920f833") from Python.)

Important: the left arm is parked in this data

Dimensions 0–5 (the left arm) hold a single constant across every frame — the recording ran with arms: 1.0, so the left arm never followed the operator. This policy therefore cannot drive the left arm, and it will not transfer to a left-arm version of the task. A bimanual task needs its own recording with arms: 0.0.

Hardware caveat

The gripper is being replaced (Fin-Ray blade 68 mm → 90 mm). Once that lands, the arm's dynamics change and this checkpoint's real-world behaviour has to be re-measured.

Not yet run on a robot

As of 2026-08-21 this checkpoint has never driven the physical arm. Any policy-to-motor path must go through the project's safety wrapper first (NaN/inf rejection, per-joint absolute and per-step clamps, Cartesian envelope rejection, base zeroing, head lock).