HABILIS Brain 0: Geometry-Change Supervision for Vision-Language-Action and Residual Flow Recovery
Paper • 2609.25558 • Published
Links
- Paper: arXiv:2609.25558
- Code: GitHub
Checkpoints
LIBERO policy — gc-vla-gcrf.ckpt
A unified GC-VLA + GCRF checkpoint for the LIBERO policy runtime. It contains the base policy, causal router, and residual components in one archive. The LIBERO action interface is single-arm and uses a 7D end-effector delta.
GC-VLA Stage-2 — gcvla_stage2.pt
A Stage-2 checkpoint for downstream adaptation, pretrained with geometry-change (GC) supervision to learn how scene geometry changes alongside robot actions.
Stage-2 action format
The canonical Stage-2 action target is a bimanual 14D end-effector delta, arranged as two 7D arm blocks:
left:[dx, dy, dz, d_rx, d_ry, d_rz, gripper]right:[dx, dy, dz, d_rx, d_ry, d_rz, gripper]
Position deltas are in metres and rotation deltas are rotation vectors in radians, expressed in the local TCP frame between adjacent action samples. The normalized gripper command uses 0 for open and 1 for closed.
The model transports this 14D target in a 32D tensor: dimensions 0–13 carry the action values; dimensions 14–31 are zero padding and masked. The padded dimensions are not additional robot actions.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support