An NVIDIA GR00T N1.6 (3B) checkpoint fine-tuned for a single-arm pick-and-place task on the SO-101 follower arm: pick up a cotton ball and place it in a bowl. Covers two object variants, each with its own language instruction:

Pick up the pink cotton ball and place it in the bowl
Pick up the yellow cotton ball and place it in the bowl

This is the original, most extensively tested checkpoint in a broader research system: it is the manipulation policy ("body") supervised by an anticipatory, Qwen3-Omni-based human-robot-interaction monitor ("brain") that watches the same camera and microphone and can stop, redirect, or confirm completion of the task in real time. The system was iterated on and validated over many live sessions on real SO-101 hardware.

Training

Base model: NVIDIA GR00T N1.6 (3B parameters)
Embodiment: SO-101 follower arm — single 6-DOF arm, parallel-jaw gripper, single camera (640x480, center-cropped to 640x360 for training/inference)
Training data: teleoperated pick-and-place demonstrations for both object variants (dataset not yet published to the Hub)
Checkpoint step: 20,000

Intended Use

Serving as the manipulation policy in a real-time control loop, optionally supervised by an external monitor that starts, stops, or redirects the policy based on inferred human intent (this checkpoint's primary use case). Not intended as a general small-object pick-and-place model — training data covers exactly two objects (differentiated by color) and one receptacle.

How to Serve

Using Isaac-GR00T's reference eval server:

bash

python gr00t/eval/run_gr00t_server.py \
    --embodiment_tag NEW_EMBODIMENT \
    --model_path tysyuvraj/groot-n16-so101-pickplace \
    --host 0.0.0.0 --port 5555 --strict

Known Behavior

Both object variants complete reliably end to end (cold-start → pick → place → confirmed completion → return to a resting pose) across many live sessions.
The yellow-ball task's completion is harder to detect visually than the pink-ball task's (lower contrast against the receptacle), which affected an earlier version of the external completion-detection system supervising this policy. This is a perception-side characteristic of the scene, not a defect in this policy's grasping or placing behavior itself.

Citation

bibtex

@misc{singh2026grootn16so101,
  author = {Singh, Yuvraj},
  title  = {GR00T N1.6 -- Cotton Ball Pick-and-Place (SO-101)},
  year   = {2026},
  url    = {https://huggingface.co/tysyuvraj/groot-n16-so101-pickplace}
}