Preflight GR00T N1.7 SO-101 — DAgger round 3

The policy behind Preflight and the reference that Autopilot improves: GR00T N1.7 for the SO-101 arm, fine-tuned in a MuJoCo twin to pick a block and place it on another, then improved by three rounds of DAgger (the scripted expert records recoveries from the policy's own failure states). Built for the Nebius × NVIDIA Physical AI hackathon. Every experiment and verdict: Preflight Experiments.

Simulation only. It has not run on a physical arm.

Results

Each suite is 96 rehearsals in the randomised twin (12 scenes × 8 perturbations, seed 42), paired scene by scene across checkpoints.

suite	v5 (scripted demos)	this checkpoint
held_out_objects (colours and sizes never seen)	52	57
seen_vocabulary	75	78
clean_many	81	88
clean_top_down_many (another camera)	83	85
real_1_pomade_tin_many (a real object as distractor)	74	87
in_path_pomade_tin_many (a real object in the arm's path)	73	88
pooled	438 / 576	483 / 576
held_out_task_basket (a task never trained)	–	6 / 96
held_out_task_ring (a task never trained)	–	0 / 96
so101-benchmark (external)	0 / 50	0–1 / 50

Promoted by a gate fixed before round 1 was scored: pooled paired gain (McNemar) over the served policy and no suite significantly worse. The new-object gain is one suite of six from one seed; new tasks fail almost every time.

Training

Base: hqfang/groot-so100_101, itself a fine-tune of nvidia/GR00T-N1.7-3B.
Data: 800 scripted twin demos (v5) plus three DAgger rounds of expert recoveries, upweighted ×8; 4,000 steps, batch 8, lr 1e-5 (train_config.json). Instruction: Pick blue cube and place on red cube.
Code: charleneleong-ai/nebius-physical-ai (scripts/dagger_round.sh, scripts/benchmark.py).

Use

LeRobot 0.6.1 GR00T N1.7 policy; vlm_processor/ is the base checkpoint's processor, needed to load it.

git clone https://github.com/charleneleong-ai/nebius-physical-ai && cd nebius-physical-ai
hf download chaleong/preflight-gr00t-so101-dagger-r3 --local-dir checkpoints/dagger-r3
python scripts/lerobot_policy_server.py checkpoints/dagger-r3 --port 5559     # in the LeRobot environment
uv run python scripts/benchmark.py run dagger-r3 --port 5559 --out bench.json

Licence

A derivative of NVIDIA GR00T N1.7, distributed under the NVIDIA License that comes with it: research and evaluation use only (non-commercial). The code in the GitHub repository is Apache-2.0; these weights are not.