XYZPIT/r1lite-ee-no_aux-150k

R1 Lite EE — GR00T N1.7, no auxiliary losses, 150k steps

Fine-tuned from NVIDIA GR00T-N1.7-3B for R1 Lite end-effector actions. This repository contains the inference weights and processor artifacts from training step 150,000. Use the Isaac-GR00T N1.7 codebase; this custom model requires its model/processor registration.

Training

• Dataset: 260413_r1lite_ee_trimmed_gr00t (50 episodes; dataset not included).
• Started from nvidia/GR00T-N1.7-3B, with a fresh optimizer.
• 150,000 steps; global batch size 16; learning rate 1e-4; cosine schedule; warmup ratio 0.05.
• AdamW; weight decay 1e-5; training seed 42; bfloat16 mixed precision.
• Frozen language and vision backbone; trainable projector, diffusion model and VLLN.
• Standard single-step flow-matching objective; trajectory consistency, action rollout and velocity smoothness weights are all 0.0.

Inputs and outputs

Embodiment tag: new_embodiment. The exact modality definition is in r1lite_ee_config.py and processor_config.json.

• Cameras: head, left wrist, right wrist RGB.
• State: left/right arm joints, grippers, and end-effector poses (xyz + rotation 6D).
• Actions: left/right grippers and end-effector poses; 20 action dimensions in total.
• Modality action window: 32 frames. The model's internal action horizon is 40.
• End-effector actions use relative representations; gripper actions use absolute representations.

Usage

Within an installed, compatible Isaac-GR00T N1.7 environment:

from huggingface_hub import snapshot_download
from gr00t.policy.gr00t_policy import Gr00tPolicy

checkpoint = snapshot_download("XYZPIT/r1lite-ee-no_aux-150k")
policy = Gr00tPolicy(
    embodiment_tag="new_embodiment",
    model_path=checkpoint,
    device="cuda",
)

The policy uses the separately hosted nvidia/Cosmos-Reason2-2B backbone configuration/processor assets as required by Isaac-GR00T.

Open-loop evaluation

Training-data evaluation: episodes 0–9, first 900 frames each, execution horizon 16, Euler with 4 denoising steps. Seed 42 is reset before each trajectory.

Model	Mean unnormalized MSE	Mean unnormalized MAE
This no-aux model	2.883978	0.079286
TCFM 150k, matched seed evaluation	2.655958	0.068620

These metrics mix gripper, position and rotation-6D components. This single-seed evaluation uses training episodes; it does not establish held-out generalization or closed-loop robot success. Per-episode results are in open_loop_metrics.json.

License and attribution

Derived from NVIDIA Isaac GR00T N1.7. The base model identifies the NVIDIA Open Model License Agreement as its license. The LICENSE file is preserved from the local base-model distribution.