ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action Models
Published in Advances in Neural Information Processing Systems (NeurIPS 2026), 2026
ROCKET aligns a VLA backbone to a 3D vision foundation model through a shared projector and a Matryoshka-style sparse activation scheme, so spatial cues at multiple depths can be distilled into the policy without modality-specific heads at inference. Combined with a training-free layer selection strategy, it reaches 98.5% average success on LIBERO using about 4% of the compute budget of prior state-of-the-art methods, with further gains on LIBERO-Plus and RoboTwin.
Accepted to NeurIPS 2026. Code and model weights: CASE-Lab-UMD/ROCKET-VLA.