Downloads · 30 days
9
33% of all-time downloads
jren313/g1-inspire-piston-starvla-n1d7-wristcam
g1-inspire-piston-starvla-n1d7-wristcam is a robotics model from jren313. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as other.
A 3-camera GR00T-N1.7 policy for the Unitree G1 (29 DoF) with Inspire hands, fine-tuned in starVLA on a long-horizon piston (pipette) manipulation task.
Downloads · 30 days
9
33% of all-time downloads
All-time downloads
27
Public
Repo size
9.5 GB
Likes
0
Public
Click a slice to open those files.
.pt9.5 GB · 100%
From the Hugging Face model README
A 3-camera GR00T-N1.7 policy for the Unitree G1 (29 DoF) with Inspire hands, fine-tuned in starVLA on a long-horizon piston (pipette) manipulation task.
The policy takes a head camera plus both wrist cameras, and predicts a 30-step chunk of absolute joint targets.
Task instruction the policy was trained/served with
pick up the piston with the right hand, inject it into the tube held by the left hand, then move it over the hole plate.
config.yaml # run config — the serving loader reads this
config.full.yaml # fully-resolved config
dataset_statistics.json # q99 normalization stats (REQUIRED at serve time)
starvla_gr00t_n1d7_piston_wristcam.yaml # the experiment recipe
run_starvla_GR00T_N1d7_piston_train.sh # the launch script
final_model/pytorch_model.pt # 8.9 GB — step 10000 (final)
Keep this directory layout. The serving path resolves config.yaml and
dataset_statistics.json two directories up from the .pt file. Moving the
weights out on their own leaves the server with no normalization stats and no
way to identify the training DataConfig.
| Base | nvidia/GR00T-N1.7-3B (official weights, not a prior starVLA run) |
| Framework | CosmosGR00TN1d7 — Cosmos-Reason2-2B VLM truncated at select_layer=16 + 32-layer AlternateVL DiT + 4-layer VL self-attention |
| Objective | Rectified flow matching (velocity prediction), 4 Euler steps at inference |
| Parameters | ~3.1 B |
| Embodiment id | 25 (new_embodiment) |
| Action horizon | 30 |
Camera and state order is a positional contract. Nothing downstream identifies a view by name — the dataloader concatenates views in list order and the Cosmos processor emits one 64-token block per image in that order. A reordered or partial set is misread silently.
Cameras (in this order):
| key | native resolution |
|---|---|
video.ego_view | 240 × 424 (head D435) |
video.left_wrist_view | 480 × 640 |
video.right_wrist_view | 480 × 640 |
Every view is resized to 224 × 224 before the backbone, so native resolution is not itself a contract — but FOV and aspect are, because the resize happens after the fact. Three views cost 192 image tokens (vs 64 for one).
State — 29 dims, flattened in this order:
state.left_arm (7), state.right_arm (7), state.left_hand (6),
state.right_hand (6), state.waist (3)
Action — 30 dims, in this order:
action.left_arm (7), action.right_arm (7), action.left_hand (6),
action.right_hand (6), action.base_height (1), action.navigate_command (3)
Actions are absolute joint targets, not deltas. The server un-normalizes them (q99) before returning, so clients receive env-space values.
Fine-tuned on birbirll/g1-inspire-piston-longhorizon-v3-wristcam
(LeRobot v2.1 — 67 episodes / 45,943 frames @ 50 fps).
| Hardware | 8 × B200 |
| Batch | 16 per device × 8 GPUs = 128 |
| Steps | 10,000 (~28 epochs) |
| LR | 2e-5 backbone / 2e-4 action head, cosine to 1e-6, 100 warmup |
| Optimizer | AdamW, β = (0.9, 0.95), wd 1e-8, grad clip 1.0 |
| Precision | bf16, gradient checkpointing on |
| Throughput | ~1.48 it/s |
Fingers pressing the plunger are a small fraction of frames but the part of the task that most often failed. This run up-weights those frames:
grasp_loss_weight: 3.0
grasp_weight_dim: 25 # right_hand[5] — the thumb channel
grasp_weight_thresh: -0.143 # normalized; == raw > 0.05, flags ~21% of frames
The action head's action_mask is already used multiplicatively and the loss
divides by action_mask.sum(), so scaling the mask on flagged frames yields an
exact weighted mean — L = Σ(w · err) / Σ(w) — not a rescaled loss.
⚠️ Do not copy grasp_weight_thresh to another dataset without checking it.
It is derived from this dataset's q01 = -0.10 / q99 = +0.25 on dim 25. On a
dataset where dim 25 is constant (q01 == q99), StateActionTransform passes
the channel through un-normalized and the threshold flags 100% of frames —
a silent no-op.
python deployment/model_server/server_policy_gr00t_zmq.py \
--ckpt_path <this_repo>/final_model/pytorch_model.pt \
--port 5555 --use_bf16
Speaks the GR00T N1.6 msgpack protocol over ZMQ, so existing GR00T clients work unchanged. On startup the server logs the camera/state/action contract it derived from the checkpoint — cross-check it against your client before trusting a rollout.
The example's data_registry/data_config.py must be importable: the checkpoint
records a data_mix name (unitree_g1_piston_wristcam_n1d7), and the server
rebuilds the training-time transform pipeline from the DataConfig registered
under it.
Derived from nvidia/GR00T-N1.7-3B and nvidia/Cosmos-Reason2-2B; the NVIDIA
license terms for those base models carry over to these weights.