IntD SO-101 Bimanual (GR00T-N1.7-3B + Intent Distillation)

Real-world bimanual SO-101 policy from "Act with Intent: Brain-to-Spine Intention Distillation" (INDI). GR00T-N1.7-3B fine-tuned with the IntD grafted intent stream ([V*;R*] distilled from a frozen Cosmos-Reason2 teacher), 60k steps (GBS 129) on 400 teleop demos (4 tasks × 100: threading, basket nesting, cross-bin stacking, drawer storage; head + left/right wrist cameras).

  • Action space: relative arm / absolute gripper joint actions (official GR00T SO-100 convention), chunk 32, executed with receding horizon 8.
  • Serve with the intd-vla fork (gr00t/eval/run_gr00t_server.py --embodiment-tag SO101_BIMANUAL, α=1).

Real-world results (n=25 per cell, 600 paired rollouts)

Condition Baseline +IntD
In-distribution clean 71 76
Held-out objects 62 67
Distractors 53 60

Largest gains on precision/long-horizon tasks (drawer storage grasp +24pp). Baseline pair: leesangoh/intd-so101-bimanual-baseline.

Downloads last month
17
Safetensors
Model size
3B params
Tensor type
BF16
·
Video Preview
loading

Model tree for leesangoh/intd-so101-bimanual

Finetuned
(140)
this model