GR00T-N1.6-3B — LIBERO All-4 Finetuned
Fine-tuned checkpoint of nvidia/GR00T-N1.6-3B on all four LIBERO task suites (Spatial, Object, Goal, LIBERO-10) with the Franka Panda embodiment.
Evaluation Results
Evaluated with 20 episodes per task across all 4 LIBERO suites (10 tasks × 20 episodes = 200 episodes per suite).
| Suite | Success Rate | Episodes |
|---|---|---|
| LIBERO-Spatial | 96.0% | 192 / 200 |
| LIBERO-Object | 100.0% | 200 / 200 |
| LIBERO-Goal | 98.0% | 196 / 200 |
| LIBERO-10 | 97.5% | 195 / 200 |
| Average | 97.9% | 783 / 800 |
Model Details
| Parameter | Value |
|---|---|
| Architecture | GR00T-N1.6 (Gr00tN1d6) |
| Vision Backbone | nvidia/Eagle-Block2A-2B-v2 |
| Backbone Embedding Dim | 2048 |
| Hidden Size | 1024 |
| Diffusion Layers | 32 |
| Diffusion Attention Heads | 32 |
| Action Horizon | 50 |
| Inference Timesteps | 4 |
| Max Action Dim | 128 |
| Model Dtype | bfloat16 |
| Embodiment | libero_panda (ID: 2) |
Training Details
| Hyperparameter | Value |
|---|---|
| Base Checkpoint | nvidia/GR00T-N1.6-3B |
| Training Steps | 40,000 |
| Global Batch Size | 16 |
| Learning Rate | 1e-4 |
| LR Scheduler | Cosine |
| Warmup Ratio | 5% |
| Weight Decay | 1e-5 |
| Tuned Modules | Diffusion model, projector, top 4 LLM layers, VLLN |
| Frozen Modules | Visual encoder, remaining LLM layers |
Usage
from gr00t.model.policy import Gr00tPolicy
policy = Gr00tPolicy(
model_path="0xAnkitSingh/GR00T-N1.6-3B_LIBERO",
embodiment_tag="LIBERO_PANDA",
use_sim_policy_wrapper=True,
)
Or serve via the inference server:
uv run python gr00t/eval/run_gr00t_server.py \
--model-path 0xAnkitSingh/GR00T-N1.6-3B_LIBERO \
--embodiment-tag LIBERO_PANDA \
--use-sim-policy-wrapper \
--port 5556