Modelpublic

StarVLA/Florence-GR00T-Bridge-RT-1

sourceHugging Faceupdated 23d agoView on Hugging Face

0likes71downloads

Model Card

Florence-2 GR00T for Bridge / RT-1

This repository contains a StarVLA QwenGR00T-framework checkpoint that uses Florence-2-large through StarVLA's Florence interface and a GR00T-style flow-matching action head. It was co-trained with the bridge_rt_1 robot-data mixture and an auxiliary VLM dataset mixture.

Model and action contract

Item	Value
StarVLA framework	`QwenGR00T` with the Florence-2 interface selected by `base_vlm`
Visual-language backbone	`microsoft/Florence-2-large`
Action head	16-layer DiT-B flow head: 768 latent width, 12 heads (64 dimensions/head); state/action decoder MLP width 1,024
Robot-data mixture	`bridge_rt_1`
Action representation	7D `delta_ee`
Configured state dimension	7
Action horizon	16
Camera input	One `image_0` RGB view, resized to 224 × 224
Inference flow steps	4
Released checkpoint	`checkpoints/steps_10000_pytorch_model.pt`

Training recipe

The packaged config.yaml records:

Setting	Value
Configured maximum steps	100,000
Save interval	10,000
VLA per-device batch	8
Auxiliary VLM per-device batch	4
Gradient accumulation	1
Warm-up	5,000 steps
Base / interface / action LR	`3e-5 / 1e-5 / 1e-4`
Training flow repeats	4 in the public run-era QwenGR00T path; YAML action-model value `8` is not read there
VLA / auxiliary VLM loss scale	`1.0 / 0.1`
`freeze_modules`	Packaged boolean `true`; the public trainer expects module paths as a string, so this value names/selects no modules
Seed	42

The saved config also retains a dinov2_vits14 field, but the selected QwenGR00T/Florence path does not construct DINO or Q-Former. It contains an environment-local Florence path. Map framework.qwenvl.base_vlm to microsoft/Florence-2-large or an equivalent local snapshot before loading.

SimplerEnv WidowX results

The following success rates are retained from the original Hub Card.

Training Steps	PutCarrotOnPlateInScene	PutEggplantInBasketScene	PutSpoonOnTableClothInScene	StackGreenCubeOnYellowCubeBakedTexInScene	Average Across Tasks
10,000	0.2083	0.8750	0.8750	0.0417	0.5000
20,000	0.4063	0.6146	0.6979	0.1354	0.4635
30,000	0.4896	0.7813	0.3854	0.1771	0.4583
40,000	0.4688	0.7500	0.3854	0.1458	0.4375
50,000	0.3750	0.5521	0.5000	0.0938	0.3802
60,000	0.3646	0.6042	0.4583	0.1042	0.3828
70,000	0.3958	0.7813	0.4688	0.1250	0.4427
80,000	0.4271	0.8125	0.5208	0.0521	0.4531
90,000	0.4063	0.5313	0.5833	0.1250	0.4115
100,000	0.4479	0.5938	0.7500	0.1042	0.4740

Only the 10k weight is present in this repository. The repository does not include per-episode evaluation logs for the table, so the 20k–100k rows cannot be reproduced from the released artifacts alone. Fluctuation across checkpoints does not by itself establish overfitting, and this Card does not recommend a deployment checkpoint.

Download and evaluation

bash

huggingface-cli download StarVLA/Florence-GR00T-Bridge-RT-1 \
  --local-dir Florence-GR00T-Bridge-RT-1

CKPT="$PWD/Florence-GR00T-Bridge-RT-1/checkpoints/steps_10000_pytorch_model.pt"
python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" --port 6678 --use_bf16 \
  --config_override framework.qwenvl.base_vlm=microsoft/Florence-2-large

Run the simulator separately using the StarVLA SimplerEnv workflow.

Evidence and limitations

—Architecture, training settings, action shape, and released step were checked against config.yaml and the Hub file tree.
—Evaluation numbers are Card-reported; raw rollout logs and weights other than 10k are not included.
—The artifact is a StarVLA framework state dict, not a standalone Transformers from_pretrained() policy.
—It is scoped to the saved Bridge/RT-1 normalization and 7D delta-EE contract. Cross-embodiment and real-robot performance are not established.
—A GPU environment compatible with the StarVLA framework and Florence-2 is required; this checkpoint has not been validated on CPU or macOS.