Control work is getting more explicit about contact, safe regions, and rollout risk

Robot manipulation work today is highly task-conditional. In the contact-aware manipulation paper, the best result comes from using torque only when contact is detected, then mixing vision and torque inside the diffusion policy. That simple timing rule matters: the method reaches 82.0% average success across three real tasks, ahead of torque gating alone at 68.0%, while vision-only and plain feature concatenation both stay at 30.0%. The error pattern is concrete. Vision-only handles empty bottles at 8/10 but drops to 0/10 on full bottles, while the adaptive fusion model keeps 9/10 on empty and 7/10 on full. This points to a narrow but important lesson for robot control: extra sensing helps when the policy knows when to trust it.

Surgical automation work is getting more explicit about where an action may safely happen. AffordTissue predicts a dense heatmap for tool-action-specific tissue interaction before contact, using text, video history, and a dedicated decoder. The payoff is in spatial accuracy, not just classification. ASSD falls to 20.557 px, compared with 60.184 px for Molmo-VLM and 81.138 px for SAM3, while PCK@0.05 reaches 0.517 versus 0.095 and 0.128. The ablations also show that the prompt content matters. Removing the language encoder raises ASSD to 43.135 px, and removing tool or action conditioning also hurts. This makes the paper useful as a safety layer story: the model can expose a predicted safe zone that a controller or human can inspect before tool contact.

World-model research today is dominated by risk framing rather than new capability gains. The key claim is that rollout-based planning can preserve and amplify small errors or attacks across imagined futures. The paper defines this with a trajectory-persistence metric and reports a 2.26x amplification ratio on a GRU-based recurrent state-space model, with a 59.5% reduction after adversarial fine-tuning. A stochastic proxy drops to 0.65x, so the exposure depends on architecture. The evidence is still limited. The DreamerV3 result shows non-zero action drift from checkpoint probing, but the paper does not provide a full downstream failure-rate benchmark. That leaves this as a useful warning and evaluation agenda, with lighter empirical grounding than the day’s two task papers.