LoongForge-GR00T-N1.7-SimplerEnv-Bridge

This directory provides a GR00T-N1.7 SimplerEnv Bridge (WidowX) SFT checkpoint trained with LoongForge.

GR00T-N1.7 is an open vision-language-action model: a Qwen3-VL / Cosmos-Reason2 backbone (nvidia/Cosmos-Reason2-2B) conditions a flow-matching diffusion transformer (DiT) action head on images, language and robot proprioception, and decodes a chunk of future actions. This checkpoint is a supervised fine-tune of nvidia/GR00T-N1.7-3B on the Bridge dataset, reproducing the official examples/SimplerEnv recipe (Bridge / WidowX) of NVIDIA Isaac-GR00T in LoongForge.

The policy emits 7D absolute end-effector actions in 8-step chunks, and evaluation executes 4 steps before replanning.

Model details

Property	Value
Benchmark	SimplerEnv Bridge (WidowX)
Base model	nvidia/GR00T-N1.7-3B
Vision-language backbone	nvidia/Cosmos-Reason2-2B (Qwen3-VL)
Embodiment tag	simpler_env_widowx
Training data	IPEC-COMMUNITY/bridge_orig_lerobot — LeRobot v2.1, 53,192 episodes, 1,893,026 frames, 5 fps, robot_type=widowx
Inputs	Task instruction plus a single exterior camera view (image_0) and robot state
Robot state	8 dimensions: EEF position 3, Euler orientation 3, a dead pad channel that is always 0 (index 6), gripper openness in [0, 1] (index 7)
Actions	7 dimensions per step (position 3, Euler 3, continuous gripper 1); absolute, not delta; chunks of 8 steps
Image preprocessing	1 image, shortest edge 256, crop_fraction 0.95, center crop 230 × 230 → 256 × 256, no letterbox
Normalization	q01/q99 percentile normalization (not mean/std) with outlier clipping, using the bundled dataset_statistics.json
Trainable modules	projector, diffusion transformer, VL-layer norms (LLM and visual tower frozen)
Weight precision	FP32, 1,031 tensors, 3,455,180,928 parameters (pytorch_model.pt, ~12.5 GB)
Training schedule	20,000 iterations, global batch 1,024 (128 × 8 GPUs, no grad accumulation), base LR 1e-4 → 0 (cosine, 1,000 warmup iterations), weight decay 1e-5 excluding bias/norm, grad clip 1.0, seed 42, bf16, DDP
State dropout	0.8 on the model side, 0.2 on the data side (matches the official --state-dropout-prob 0.8 mapping)
Reported result	63.5 % mean success over the 7 WidowX tasks (100 episodes per task, 3 runs averaged); NVIDIA N1.7 reference 62.3 %

Repository contents

pytorch_model.pt          # model weights (FP32, single-file state dict)
processor_config.json     # eval-time image / normalization spec
dataset_statistics.json   # state and action normalization statistics (q01/q99)
config.json               # descriptive model, preprocessing and training metadata
resume_meta.json          # completed steps, epoch, checkpoint format, world size

config.json is descriptive metadata only. It is not a Transformers AutoModel configuration — LoongForge still loads this checkpoint through its own explicit evaluation configuration.

Usage

Serve this checkpoint in SimplerEnv with the LoongForge evaluation stack. The entry point is the SimplerEnv config tree, examples/embodied/groot_n1_7/eval/configs/simplerenv in the LoongForge source tree. Point it at this directory:

model:
  backend: loongforge
  model_type: Gr00tN1d7
  model_name: /path/to/Cosmos-Reason2-2B

server:
  ckpt_path: /path/to/this/directory          # contains pytorch_model.pt
  dataset_statistics_path: /path/to/this/directory/dataset_statistics.json
  embodiment_tag: simpler_env_widowx
  chunk_execute_steps: 4                      # official --n_action_steps
  use_bf16: true

Acknowledgements

Built on LoongForge, SimplerEnv and the Bridge dataset.