You can deploy this model and start earning money today!

jren313/
g1-inspire-pipette-starvla-n1d7-3view-p12

Name: g1-inspire-pipette-starvla-n1d7-3view-p12
Availability: InStock

robotics

vision-language-action

gr00t

unitree-g1

starvla

jren313/g1-pipette-3view-20260819

nvidia/GR00T-N1.7-3B

finetune

Views

No views yet

G1 + Inspire · pipette-tip attachment · starVLA GR00T-N1.7 · 3 cameras (P1+P2)

A GR00T-N1.7 fine-tune that drives a Unitree G1 with Inspire hands to pick a pipette tip out of a rack and seat it on the pipette. Trained from three cameras at mixed resolution, on episodes truncated at the end of phase 2.

This policy only approaches and inserts. Its training episodes stop the frame the tip seats in the well, so it has never seen the lift, the carry, or the ejector press. On the robot it reaches the rack, inserts, and has no continuation — that is the design of the run, not a defect, but an operator waiting for an eject will wait forever.

Inputs

view	streamed	model input	tokens
`rgb` (ego)	1280x720	256x256	64
`wrist_left`	1920x1080, centre-cropped to 1080x1080	448x448	196
`wrist_right`	1920x1080, centre-cropped to 1080x1080	448x448	196

Neither side camera is wrist-mounted despite the names, which are the dataset's — measured global pan across each recording is 0.09–0.5 px where a wrist mount would show tens. They are fixed cameras beside the rack.

The two side views run at 448 rather than 256 because at 256 the green tip measures only 1–3 px in them (it is 6x10 px in the ego view); at 448 it measures 6x11 and 7x4 and the rack wells resolve individually.

State: 17-dim (waist 3 | left_arm 7 | right_arm 7), 4 frames of history at indices [-36, -24, -12, 0] (0.6 s @ 60 Hz).

Language: one constant sentence — Attach a green pipette tip from the tip rack to the pipette. The tip is green in this capture. There is no phase column; do not serve this through a phase-conditioned lane.

Outputs

Serving

Serve WITH state. The state_encoder is built for 132 * 4; passing --no_state zero-fills all four history slots.
Run the robot at action_hz 60 — the policy is velocity-conditioned.
Do not pre-downscale the frames. Send them at capture resolution and let the server resize, reading obs_image_size from config.yaml. Training resizes once with PIL from the full-resolution stored frame; a resize elsewhere substitutes a different resampler and hands the head an image the training pipeline never produced.

config.yaml and dataset_statistics.json sit two directories above the checkpoint, which is where the starVLA loader looks (share_tools.read_mode_config resolves them off checkpoint_pt.parents[1]). Keep the layout intact:

config.yaml
dataset_statistics.json
checkpoints/steps_8000_pytorch_model.pt

Training


base	`nvidia/GR00T-N1.7-3B` (Cosmos-Reason2-2B truncated at `select_layer=16` + 32-layer flow-matching DiT)
params	3.1B
`embodiment_id`	25
data	53 episodes, 37,943 frames — the phase-1+2 truncation of 69,148
steps	10,000 @ global batch 128 (8 x B200), 33.7 epochs
optimiser	`paged_adamw_8bit`, lr 2e-5 base / 2e-4 action head, cosine with min_lr
`state_dropout_prob`	0.3

This checkpoint is step 8000, not final_model. Step 8000 scored mse_score 0.000663 against step 10000's 0.000715.

Limitations, stated plainly

The score above is not a held-out number. This recipe has no eval split — every episode trains — so mse_score is measured on data the model is fitting, and it swings by a factor of 2 between adjacent log points. Judge this checkpoint on a robot, not on that number.
Perceptual aliasing is real on this task. On look-alike frame pairs, 68% have opposed next-chunk arm directions (median cosine −0.31). The 4-frame state history exists to break that tie; a memoryless serving path passes open-loop and stalls closed-loop.
One rig, one tip colour, one rack position. Nothing here has been tested for generalisation beyond the capture setup.

License

Inherits the terms of nvidia/GR00T-N1.7-3B. Check NVIDIA's license before redistributing or using commercially.

jren313/g1-inspire-pipette-starvla-n1d7-3view-p12