🦾 GigaBrain-0.7 · 3.5B vision-language-action model

GigaBrain-0.7 is an embodied foundation model with a three-system architecture. Give it three camera views, a natural-language instruction and the robot's current joint angles, and it predicts the next 50-step / 14-DoF action chunk (≈1.7 s of bimanual motion at 30 Hz) with its flow-matching action expert.

Configured for the released AgileX Cobot Magic, fixed base deployment profile (embodiment 6, agilex_cobot_magic, 14-D state → 14-D action). There is no robot attached here, so the chunk is plotted rather than executed.

The joint state is what the robot reports back; for your own images you can keep the default pose or paste a real 14-D reading.

Real AgileX Cobot Magic frames from the official GigaBrain-0.7 sample data

Notes

  • The checkpoint is the pre-trained base model, not a task-specific post-trained policy, so the example frames are the fairest way to read its behaviour: the dashed curves are the human teleoperation actions actually recorded at that frame.
  • Preprocessing follows the authors' deployment path exactly — float_native RGB at native resolution, resize-with-pad to 224×224, SigLIP [-1, 1] normalisation, quantile normalisation of the state, delta actions for the joints and absolute targets for the grippers.
  • Normalisation statistics were recomputed with the repo's compute_norm_stats_fast.py recipe over the 1 012 released episodes of open-gigaai/GigaBrain-0.7-SampleData (Apache-2.0), which is also where the example frames and joint states come from.
  • Code: open-gigaai/giga-brain-0 · giga-models (both Apache-2.0; giga_models v1.1.0 is vendored into this Space).