LoongForge-GR00T-N1.7-LIBERO
These directories provide the four GR00T-N1.7 LIBERO SFT checkpoints trained
with LoongForge β one independent
fine-tune per suite (libero_spatial, libero_object, libero_goal,
libero_10), as in the official
nvidia/GR00T-N1.7-LIBERO
release.
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. Each checkpoint here is a supervised fine-tune
of nvidia/GR00T-N1.7-3B on one LIBERO suite, reproducing the official
examples/LIBERO recipe of NVIDIA Isaac-GR00T
in LoongForge.
The policy emits 7D delta end-effector commands in 16-step chunks, and
evaluation executes 8 steps before replanning (official --n_action_steps 8).
Checkpoints
| Suite | Checkpoint | Training data |
|---|---|---|
| libero_spatial | libero_spatial/ |
432 episodes, 52,970 frames |
| libero_object | libero_object/ |
454 episodes, 66,984 frames |
| libero_goal | libero_goal/ |
428 episodes, 52,042 frames |
| libero_10 | libero_10/ |
379 episodes, 101,469 frames |
Model details
| Property | Value |
|---|---|
| Benchmark | LIBERO, 4 suites Γ 10 tasks |
| Base model | nvidia/GR00T-N1.7-3B |
| Vision-language backbone | nvidia/Cosmos-Reason2-2B (Qwen3-VL) |
| Embodiment tag | libero_sim |
| Training data | the four *_no_noops_1.0.0_lerobot datasets above β LeRobot v2.1, 20 fps, robot_type=franka; the no_noops variant drops the LIBERO settle steps |
| Inputs | Task instruction, an exterior camera view (image) and a wrist camera view (wrist_image), plus robot state |
| Robot state | 8 dimensions: EEF position 3, orientation as axis-angle 3 (quat2axisangle of robot0_eef_quat), and the 2D gripper finger joints robot0_gripper_qpos (2). Note the gripper occupies two slots, unlike the 1D openness of the SimplerEnv embodiments |
| Actions | 7 dimensions per step (translation delta 3, rotation delta as axis-angle 3, gripper 1); delta end-effector commands in the OSC_POSE action space; chunks of 16 steps |
| Image preprocessing | 2 images, shortest edge 256, crop_fraction 0.95, center crop 230 Γ 230 β 256 Γ 256, no letterbox; both views are rotated 180Β° inside the LIBERO adapter, as the official client does |
| Normalization | q01/q99 percentile normalization (not mean/std) with outlier clipping, using each checkpoint's 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 per checkpoint (pytorch_model.pt, ~12.5 GB) |
| Training schedule | 20,000 iterations, global batch 640 (80 Γ 8 GPUs, no grad accumulation), base LR 1e-4 β 0 (cosine, 1,000 warmup iterations), weight decay 1e-5 excluding bias/norm, grad clip 1.0, seed 42, bf16, DDP; TEFusedAdamW for spatial/object/10, plain AdamW for goal |
| State dropout | 0.2 on both the data side and the model side (the official --state-dropout-prob 0.2) |
Repository contents
The same layout in every suite directory:
libero_spatial/
βββ pytorch_model.pt # model weights (FP32, single-file state dict)
βββ processor_config.json # eval-time image / normalization spec
βββ dataset_statistics.json # state and action normalization statistics (q01/q99)
βββ config.json # descriptive model, preprocessing and training metadata
βββ resume_meta.json # completed steps, epoch, checkpoint format, world size
libero_object/ # same five files
libero_goal/ # same five files
libero_10/ # same five files
config.json is descriptive metadata only. It is not a Transformers
AutoModel configuration β LoongForge still loads these checkpoints through its
own explicit evaluation configuration.
Usage
Serve a suite checkpoint in LIBERO with the LoongForge evaluation stack. The
entry point is the LIBERO config tree,
examples/embodied/groot_n1_7/eval/configs/libero
in the LoongForge source tree (libero_<suite>.yaml). Point it at the matching
directory:
benchmark:
name: libero
suite: libero_spatial # libero_spatial | libero_object | libero_goal | libero_10
max_steps: 720 # official --max_episode_steps
num_steps_wait: 0
continuous_gripper: false # official normalize_gripper_action(binarize=True)
control_mode: delta
model:
backend: loongforge
model_type: Gr00tN1d7
model_name: /path/to/Cosmos-Reason2-2B
state_encoding: libero_ee_axis_angle
action_encoding: axis_angle
server:
ckpt_path: /path/to/libero_spatial # the suite directory, contains pytorch_model.pt
dataset_statistics_path: /path/to/libero_spatial/dataset_statistics.json
embodiment_tag: libero_sim
chunk_execute_steps: 8 # official --n_action_steps
use_bf16: true
Acknowledgements
Built on LoongForge, LIBERO and the LIBERO LeRobot datasets.
Model tree for LoongForge/LoongForge-GR00T-N1.7-LIBERO
Base model
nvidia/GR00T-N1.7-3B