Interesting take, i wonder if we would ever get to the point of being able to guarantee safety in human interaction with robots say in the next 15 years.
I think it's going to be a combination of relaxing our standards and targeting different tasks. It's not that these problems are insolvable - they are absolutely solvable. But they are not the low-hanging fruit. I've also always found it interesting how we have different standards for safety in different contexts. For example, the safety system for a subway train coming into the station on the T in Boston is a big yellow strip.
I wonder if all of our people-interacting robots will slowly get covered in more squishy and sensorized material to give it an inherent safeness (or maybe inflatable robots will be more possible). But I agree it feels like there’s always going to be a usefulness/danger coupling that is going to be hard to overcome. Do you think there’s a way for robots interacting with people physically to know if certain strategies for doing a task have a higher allowable force limits without being unsafe (like an experienced person could maybe wrap a blood pressure monitor with less force?)
Or can robots better use mechanics and contact of the interaction if they had high enough bandwidth to allow for higher force limits (but not sure what it would even mean to have an ISO standard for that)
Baymax was inspired by soft-robotics work at CMU, where inflatable arms were attractive because compliance lowers peak force and spreads contact area, which drops the pressure at any single point. The catch is that softness usually costs you precision and payload, so the usefulness/danger coupling tends to come back as a usefulness/controllability coupling.
Task space controllers let us bound force along specific directions or phases of a task, so a robot can be stiff where it needs to work the cuff and compliant everywhere else. The hard part is that hand-modeling those safe envelopes for every task we care about is a lot of work, and it gets harder once the behavior comes from a VLA, where the path from a high-dimensional policy to a certifiable force bound is murky.
The ISO co-bot limits (TS 15066) are written for unexpected contact, where the robot is blind to where the person is and has to promise safety everywhere. Intended, sensed contact is a different regime. If the robot knows it is holding your arm and can regulate force at high bandwidth how tightly can the robot control known contact. The everywhere-limit stops being the right model. TS 15066 already separates quasi-static from transient contact, so I can imagine a future standard that certifies a sensing-and-control loop and its response bandwidth. Building methods that let us trust a learned policy inside that loop is the open problem, and an important one at that.
Interesting take, i wonder if we would ever get to the point of being able to guarantee safety in human interaction with robots say in the next 15 years.
I think it's going to be a combination of relaxing our standards and targeting different tasks. It's not that these problems are insolvable - they are absolutely solvable. But they are not the low-hanging fruit. I've also always found it interesting how we have different standards for safety in different contexts. For example, the safety system for a subway train coming into the station on the T in Boston is a big yellow strip.
Thanks for the fun post!
I wonder if all of our people-interacting robots will slowly get covered in more squishy and sensorized material to give it an inherent safeness (or maybe inflatable robots will be more possible). But I agree it feels like there’s always going to be a usefulness/danger coupling that is going to be hard to overcome. Do you think there’s a way for robots interacting with people physically to know if certain strategies for doing a task have a higher allowable force limits without being unsafe (like an experienced person could maybe wrap a blood pressure monitor with less force?)
Or can robots better use mechanics and contact of the interaction if they had high enough bandwidth to allow for higher force limits (but not sure what it would even mean to have an ISO standard for that)
Baymax was inspired by soft-robotics work at CMU, where inflatable arms were attractive because compliance lowers peak force and spreads contact area, which drops the pressure at any single point. The catch is that softness usually costs you precision and payload, so the usefulness/danger coupling tends to come back as a usefulness/controllability coupling.
Task space controllers let us bound force along specific directions or phases of a task, so a robot can be stiff where it needs to work the cuff and compliant everywhere else. The hard part is that hand-modeling those safe envelopes for every task we care about is a lot of work, and it gets harder once the behavior comes from a VLA, where the path from a high-dimensional policy to a certifiable force bound is murky.
The ISO co-bot limits (TS 15066) are written for unexpected contact, where the robot is blind to where the person is and has to promise safety everywhere. Intended, sensed contact is a different regime. If the robot knows it is holding your arm and can regulate force at high bandwidth how tightly can the robot control known contact. The everywhere-limit stops being the right model. TS 15066 already separates quasi-static from transient contact, so I can imagine a future standard that certifies a sensing-and-control loop and its response bandwidth. Building methods that let us trust a learned policy inside that loop is the open problem, and an important one at that.