kuma-levi/gr00t-n1.7-test

GR00T N1.7 fine-tuned: Unitree G1 two-handed pick, carry and place

The robot picks a bottle up between two open palms, turns at the waist, sets it down beside where it started, and withdraws. Fine-tuned from nvidia/GR00T-N1.7-3B on demonstrations generated in NVIDIA Isaac Sim.

Pipeline, dataset generator and evaluator: https://rtl.acryl.ai/aaai/physical-ai

What it commands

Fifteen numbers, as relative joint targets over a sixteen-step horizon:

Slice	Joints
right_arm	shoulder pitch, roll, yaw, elbow, wrist roll, pitch, yaw
left_arm	the same seven
waist	waist yaw

Observations are two 224x224 RGB images - a fixed scene camera and a wrist camera - plus the fifteen measured joint angles and a language instruction. Embodiment tag new_embodiment.

The hands are never commanded. They are five-finger props held open, and the bottle rides between two opposed palms.

Training data

episodes	80, none discarded
frames	13,518
recorded at	10 Hz
task	pick up the pill bottle and set it down to the side

The demonstrations are scripted, not teleoperated. What makes them worth anything is the physics they were recorded under:

• gravity on, and actually acting. An earlier version of the generator zeroed the articulation's velocity every physics step to stop the base drifting, which made the arms weightless whatever the gravity flags said.
• driven, never teleported. The arm follows joint position targets. The state stored beside each image is what the encoders read, not what the joints were told, so it carries the real tracking error.
• gains a real arm could use. The asset's own drive gains demand 53.6 N.m from a joint rated for 25. At kp 2000 / kd 60 the same motion costs 21.5 N.m.

Measured over the whole dataset: tracking error 0.15 degrees mean and 1.01 at the 99th percentile, peak shoulder torque 21.5 N.m against a 25 N.m limit, and no arm link inside the torso in any frame.

What this model cannot do

It cannot grasp. The bottle is attached between the palms kinematically. No training frame contains finger contact, so nothing here transfers a grasping skill to hardware.

It has seen one room, one lighting setup, one bottle and one camera pose. There is no domain randomisation. Real camera images are out of distribution.

The simulated pelvis is held in place. A real G1 standing on its legs moves its torso to stay balanced, so the hands will not land where this predicts.

Using it

Served over ZMQ by gr00t.policy.server_client.PolicyServer; the repository above has the client, the modality config and the closed-loop evaluator.