XYZPIT/vlash-random4-gr00t-n1.6-160000

VLASH GR00T N1.6 — R1 Lite (random delay 0–4)

Finetune of Isaac GR00T N1.6 for the Galaxea R1 Lite embodiment, trained with VLASH temporal delay augmentation: for every sample a delay k ~ U[0, 4] is drawn and the state/action windows are shifted by k while video and language stay at the base timestep.

video/language:  t                                   (stale observation)
state:           t + k
action chunk:    action[t + k : t + k + horizon]

The policy therefore predicts an action chunk that begins k steps after the observation it was given, which is what makes asynchronous chunk execution possible on the real robot: the next chunk can be computed while the current one is still executing.

Details

Base model	Isaac GR00T N1.6
Checkpoint	step 160000
Embodiment tag	new_embodiment (R1 Lite, bimanual + 2 grippers)
Action horizon	16
max_delay_steps	4 (random per sample)
use_state_ground_truth	true (delayed state is the measured state at t + k)
Tuned layers	backbone LLM layers 12–15 + action head

Open-loop evaluation

Delay sweep with gr00t/eval/delayed_open_loop_eval.py, 3 trajectories × 200 steps, horizon 16. Unnormalized action error, averaged over trajectories.

Arm joints only (rad):

delay	0	1	2	3	4	5	6	7	8
MSE	.00103	.00124	.00097	.00049	.00037	.00034	.00036	.00048	.00048
MAE	.01344	.01358	.01227	.01129	.01022	.01033	.01071	.01181	.01196

All keys concatenated (dominated by the grippers, which span 0–100):

delay	0	1	2	3	4	5	6	7	8
MSE	.00605	.00619	.00626	.00561	.00586	.00551	.00571	.00599	.00629
MAE	.02768	.02792	.02728	.02535	.02599	.02520	.02588	.02649	.02739

Arm error bottoms out around delay 4–6 — roughly 3× lower than evaluating at delay 0 — matching the 0–4 training distribution, and degrades again at delay 7–8 (outside the trained range). Gripper error is flat across delays: those are discrete open/close transitions that a few steps of shift do not fix.

Caveat: this sweep was run on the training dataset, so the numbers are train-set fit, not held-out generalization.

Files

Inference artifacts only — optimizer.pt, scheduler.pt, rng_state_*.pth, and the trainer state are not included, so this checkpoint cannot be used to resume training.

Usage

from gr00t.policy.gr00t_policy import Gr00tPolicy
from gr00t.data.embodiment_tags import EmbodimentTag

policy = Gr00tPolicy(
    embodiment_tag=EmbodimentTag.NEW_EMBODIMENT,
    model_path="XYZPIT/vlash-random4-gr00t-n1.6-160000",
    device="cuda",
)
action_chunk, _ = policy.get_action(observation)

Because of the delay augmentation, the returned chunk is meant to start executing a few steps after the observation timestamp rather than immediately.