Contact-time control and VLA safety define the day’s robotics papers

Move-Then-Operate argues that high-precision manipulation improves when the policy treats approach motion and contact work as separate behaviors. Its dual-expert design uses one expert for moving and one for operating, with a router selecting the phase per action chunk. On 8 RoboTwin2 tasks with 50 demonstrations per task, it reports 68.88% average success, above pi_0 at 44.75%, RDT at 35.63%, and ACT at 31.63%. The gains are large on contact-heavy tasks such as Click Bell at 99% versus 44% for pi_0 and Place Cans Plasticbox at 79% versus 34%. The paper also claims peak performance in 40% fewer training steps.

Tube Diffusion Policy focuses on contact-rich control under uncertainty. It keeps diffusion for chunk-start action generation, then adds a learned streaming feedback flow so the robot can correct actions at each step from fresh visual and tactile input. The paper reports consistent gains over prior imitation-learning baselines on Push-T and three additional visual-tactile tasks, plus two real-world experiments with stronger reactivity under disturbances. The excerpt does not include exact success rates, so the main grounded takeaway is architectural: reactive within-chunk correction is being treated as a core control requirement for tactile manipulation.

Safety is getting treated as a first-class research problem for vision-language-action models, not a side note attached to model scale. The VLA safety survey lays out a concrete threat map across training-time and inference-time attacks, and across training-time and runtime defenses. It covers threats such as data poisoning, backdoors, adversarial patches, cross-modal perturbations, jailbreaks, and freezing attacks, then ties them to evaluation and deployment issues across six domains. The paper is a survey rather than a new benchmark report, but it is useful as a sign of where the field now needs shared evaluation and runtime protection.