Contact-Aware VLA Manipulation

Robot teams running VLA policies on contact-rich manipulation should split the control loop by sensor timing. DAM-VLA encodes language once, updates vision sparsely, keeps dense force and proprioception histories at the control rate, and lets the action head read the latest buffered latents at every step. This is a concrete retrofit for labs where synchronous VLA inference stalls on slow camera or language inputs while force spikes arrive faster than the policy can react.

The practical test is a narrow A/B on one hard contact task, such as socket insertion or button pressing: compare the current synchronous policy against a buffered version that preserves the existing VLA weights where possible and adds gated cross-attention for fast modalities. DAM-VLA reports 95.2% average success across seven real Franka tasks, compared with 40.95% for the strongest synchronous baseline, while sustaining 100 Hz control. The same summary reports that naive high-frequency synchronous X-VLA_100 falls to 21.9%, so the check should measure success, action latency, and whether camera upsampling degrades behavior.

Bimanual manipulation teams should add a tactile correction stage when failures come from hidden misalignment, pressure, blockage, or contact on the wrong surface. TacCoRL starts with a pretrained VLA, adds tactile tokens over a recent history window, gates out background touch readings, then trains contact recovery in a real-aligned simulator with a supervised anchor on real trajectories. This fits insertion, assembly, and puzzle-placement workflows where collecting many near-failure hardware rollouts is slow and can damage sensors.

A small adoption test can start with one task and two policies: the current vision-only VLA after RL post-training, and the same policy with tactile tokens plus the contact gate. TacCoRL reports 72.5% average success across four real bimanual contact-rich tasks, compared with 50.0% for the vision-only RL post-trained policy. The simulator step matters: direct sparse-reward RL from the base VLA gets 0.0 success across the four simulated tasks, while co-training and a real-data anchor improve transfer.

Teams evaluating VLA policies under changed camera pose, lighting, object geometry, deformable state, or contact tolerance should test a frozen world-action model as an added prior. World Pilot keeps the world-action model frozen, injects a scene-evolution latent into the perception stream, and feeds an anticipated action trajectory to the action generator. This is a concrete evaluation path for labs that already have a trained VLA and see success drop when the same task is run under new visual or physical conditions.

A first check is to cache world-action model outputs during training, then run the model online at each decision step for a fixed OOD suite. World Pilot reports 84.7% total success on LIBERO-Plus zero-shot OOD, ahead of ABot-M0 at 80.5% and Cosmos Policy at 79.7%. In real-robot tests across four tasks and twelve settings, it reports the highest success rate in every listed table cell, including Container-Lid Alignment with lid-pose OOD at 65% versus 15% for the best listed baseline.