ACT · dvla-can 250 Hz (delta + trim) — checkpoint 50k
Best-performing 250 Hz policy on the dvla-can dynamic place task (rolling can → bowl,
intercept-while-moving). ACT trained on
mickeykang/dvla-can-250hz-events-250fps-delta-trim
(s20 scene distribution, 250 Hz control, RGB + DVS event columns).
Results (clean in-domain scene set, n=20)
| checkpoint | success |
|---|---|
| 50k (this repo) | 12/20 = 60% |
| 100k/150k/200k | worse (long-training collapse measured) |
- Repeat variance on the same checkpoint is ±2–3/20 (n=20 eval) — treat 60% as a point estimate.
- Scene set = the corrected in-domain 20 scenes (post contamination fix; membership-checked
against the training
meta/camera.jsonl).
Contents
050000/pretrained_model/ # lerobot ACT weights + config (inference)
050000/training_state/ # optimizer state (resume)
Eval protocol (matters — measured pitfalls)
- Scene flags OFF: this model predates the mo5 textured scene. Evaluate on the ORIGINAL
plain scene (
BASKET_SCENE/TEXTURED_OBJS/CAN_TEXTUREall unset) or renders will not match training pixels (policies are render-sensitive: 5% vs 33% across GL backends). - Eval server on OSMesa with 240 s episode timeout (EGL aborts when GPUs are busy).
- control_freq 250, delta actions, trim-hold preprocessing — as in the dataset name.
- Loader: the DynamicVLA_mujoco patched lerobot (temporal-fusion + n_obs guard patches);
see
github.com/mickeykang16/DynamicVLAbranchmujoco.
Provenance
Training/eval code: github.com/mickeykang16/DynamicVLA (branch mujoco).
Run name: act_tf250_s20_dlt_tr. Uploaded 2026-07-23.