Researchers have built a safety system for robots that skips per-robot retraining.
The project, called CrossSafe, tackles a real gap in robotics: a movement that's safe for one robot's arm can be a collision for another with different proportions. Rather than hand-coding safety rules for each machine, the researchers trained a single value function using Hamilton-Jacobi reachability analysis, a mathematical framework for predicting worst-case outcomes, in a latent space that encodes each robot's morphology and kinematics. They tested it across five bimanual robot embodiments and five manipulation tasks with whole-body collision-avoidance constraints. The policy was trained jointly across all five tasks and four of the five robots, then applied without further training to the fifth, holdout robot.
That holdout test is the actual finding. The policy lowered the collision rate on a robot it had never seen, and the researchers found accuracy improved as more embodiments were added to training. That points toward a future where safety systems, not just manipulation skills, could be a reusable layer across a fleet of different robots rather than a bespoke build for each one.
It's an early academic result, not a shipping product: five embodiments in a lab is a long way from the mess of real warehouses and homes. But as generalist robot policies get more attention, betting that safety reasoning is transferable is a more defensible wager than assuming every robot needs its own safety net from scratch.