Modelpublic

seungkukim/rldx1-pt-real6k-syn10k-pcpass-60k

sourceHugging Faceotherupdated 11d agoView on Hugging Face

0likes16downloads

Model Card

RoboCurate V2 — RLDX-1 PT, real6k + Pre-Contact-filtered syn10k (60K, WSD)

Pre-training checkpoint (step 60,000) of an RLDX-1 vision-language-action model, trained on real robot data mixed 1:1 with generated clips that were curated by the Pre-Contact filter — RoboCurate V2's simulator-free, action-level verification of synthetic robot video.

This is the filtered arm of a controlled comparison: it shares its first 45,000 steps with an unfiltered control, and the two differ only in which synthetic clips the final decay leg sees.

Training

Warmup–Stable–Decay (WSD) over 60,000 steps on 4× H100:

phase	steps	synthetic half	learning rate
warmup	0 → 3,000	unfiltered 10,000 clips	0 → 1e-4
stable	3,000 → 45,000	unfiltered 10,000 clips	1e-4
decay	45,000 → 60,000	Pre-Contact PASS, 5,775 clips	1e-4 → 0

Because the branch point at 45K is shared, any difference against the control is attributable to the filter alone, not to the optimisation path.


base model	RLDX-1 (`RLDX-1-PT-IMG-noAH`, action-head-stripped)
global batch / accum	256 / 1 (64 per GPU × 4)
optimizer	AdamW, bf16, seed 42
LR schedule	`constant_with_warmup` + WSD decay (`decay_start_step=45000`, `decay_end_step=60000`)
action horizon	16
video length / cognition tokens	1 / 64
state dropout	0.0
final train loss	0.0044

Data

half	embodiment tag	store	episodes	frames
real	`GENERAL_EMBODIMENT`	real6k @ 16 fps	6,551	1,258,734
synthetic (0→45K)	`NEURAL_GR1`	syn10k, unfiltered	10,000	930,000
synthetic (45→60K)	`NEURAL_GR1`	syn10k, Pre-Contact PASS	5,775	537,075

Sampled 1:1 by mix_ratio (a per-spec probability, independent of store size), so the decay leg draws the same number of synthetic samples from the smaller filtered pool — 3.57 epochs over its 537,075 sampleable steps.

The Pre-Contact filter keeps 57.75% of the generated clips. It verifies, without a simulator, that the predicted actions bring the hand into a physically consistent pre-contact relationship with the target object, using forward kinematics over a rollout of the predicted action chunk plus metric depth and object extent.

Intended use

Research on synthetic-data curation for robot learning: as a pre-training initialisation for downstream fine-tuning, and as the treatment arm against its unfiltered control. It is not an instruction-following policy ready for deployment on hardware.

Limitations

—The synthetic half normalises with base-checkpoint statistics rather than the synthetic store's own — deliberate, so the arms stay comparable, but it means the generated data is not normalised to itself.
—Generated clips carry actions inferred by an inverse dynamics model, not measured ones; observation states in that half are zero.
—Real data is GR-1 humanoid tabletop manipulation. Nothing here transfers to another embodiment without re-tagging and re-normalising.

Provenance

Part of RoboCurate V2, stage 7 (VLA training). Derived from RLDX-1; redistribution is subject to whatever terms govern the base model.