Robot VLA papers demand contact-aware control and behavior-level validation

DyGRO-VLA treats reinforcement learning as a controlled edit to a base robot policy. It freezes the base VLA and trains routed residual experts that add delta action chunks. On LIBERO, it reports 97.1% average success, a +4.4 point gain over its offline base, with a +9.8 point gain on LIBERO-Long.

AffordVLA attacks a different failure mode: the policy may choose the right object and still touch the wrong part. It aligns intermediate VLA visual tokens with a frozen affordance teacher during training, then removes the teacher at inference. The paper reports RoboTwin gains of 20.5% in Easy and 12.8% in Hard over the previous best baseline.

RoboFlow4D uses predicted 3D gripper flows as an explicit plan that a policy can track. Its slow-fast loop plans lower-frequency trajectories and lets the action policy execute higher-frequency chunks. The paper reports LIBERO gains of +6.2 points with Diffusion Policy and +4.0 points with a DiT policy, plus under-one-second planning latency.

Visual Sculpting shows the same emphasis in deformable manipulation. It plans over dense 512×512 depth maps and spatial depth gradients, then replans with model predictive control after small batches of actions. The system produced long-horizon clay sequences with more than 100 actions, and its visual loss improved foam and dough deformation metrics in the reported held-out tests.

Event-Grounded Sparse Autoencoders adapts sparse autoencoders (SAEs) to robot policies by anchoring features to recurring rollout events. The pipeline extracts end-effector keyframes, clusters them by visual and robot-state cues, then tests features through closed-loop interventions. On OpenVLA layer 31, zeroing event-aligned features reduced success from 70.0% to 48.8%, a larger effect than window-mean, task-mean, or random-alive rankings.

The driving-safety paper tests whether Chain-of-Causation (CoC) explanations match scenes and trajectories in Alpamayo-R1-10B. Overall reasoning fidelity is 42.5% across inferences with obstacle context. The study also reports 94 missed pedestrians, 53.3% low reasoning-action consistency, and 37.9% of claimed stop cases continuing instead.