Robot VLA research centers on deployment-time adaptation and control

ICWM targets a concrete failure mode: camera view, mounting offset, and robot geometry can change after training. The robot runs a short safe probing phase, records start image, action, and end image clips, then prepends those clips as context for the policy. On LIBERO cross-view tests, it improves average out-of-distribution success by 13.0 percentage points over Multi-View BC and by 9.5 points over an explicit camera-angle baseline. The ablations are strong evidence for the mechanism: removing images drops average success by 56.4 points, and false context performs worse than no context.

Two VLA papers turn online interaction into safer policy improvement. FORCE calibrates a critic before actor updates, keeps expert and rollout buffers, and distills toward critic-preferred actions. It reports 82.3% average success on six ManiSkill tasks with an Octo backbone and 98.3% average success on six real Franka tasks after online fine-tuning, up from 45.0% behavior cloning. ROAD-VLA takes a token-level route: it shifts sampled action-token logits with calibrated advantage estimates and trains the student toward that nearby teacher distribution. Its excerpt gives task coverage and method details, but no success-rate table.

Several papers treat the action side of VLA models as a bottleneck with its own data and runtime needs. Learning Action Priors pretrains the action module on state-action trajectories before VLA training, then reuses the decoder as the action head. The reported evaluation spans 13 cross-embodiment tasks across LIBERO, RoboCasa, and a real Franka platform, though the excerpt does not include exact success rates. ACNet addresses inference delay during chunked control. It conditions the next action chunk on the motion already executed during delay and trains only about 20% of parameters on Kinetix. On Meta-World MT50, it matches full delay-conditioned retraining at 0.74 average success while reporting 91 ms latency and 11.0 Hz control frequency.

The broader framing is also becoming more action-centered. World Action Models defines predictive-action models by whether their forecasts are useful for control, and organizes methods by rendered futures, latent futures, or action reasoning without video generation. WOLF-VLA applies VLA training to whole-body humanoid locomotion with optimal-control-generated demonstrations. Its dataset contains 277 hours across six locomotion task families and 15,276 episodes on an RH5 humanoid in MuJoCo. The paper provides dataset scale and training details, but the excerpt does not provide a numeric success-rate table.