Vision-Language-Action (VLA) models that directly regress joint-space actions avoid inverse kinematics at deployment, but their training objectives can be misaligned with task success. The GRASP paper proposes injecting two priors during training to address this gap for fine-grained bimanual manipulation.
First, a physics-grounded flow-matching loss maps the model's clean-action estimate through differentiable forward kinematics at every noise level, supervising the resulting Tool Center Point pose and inter-arm relative pose. A bounded noise-level weighting compensates gradient decay. Second, the model gains a global prediction query group distilled from a frozen DINOv3 teacher via future-frame feature prediction, plus a local interaction query supervised by Grounding DINO at the attention level.
At inference, both teachers are removed, so deployment remains a standard joint-space flow policy with only the learned query pathway added. In experiments on long-horizon bimanual tasks, the two priors are complementary: the full model excels in in-distribution deployment, while the distilled query branch is preferable under strong visual randomization.