3 ms·

"You also might have better luck running dynamics simulations with PhysX than with ODE." I worked on the simulator at Anybots, and I'm pretty much convinced a d

"You also might have better luck running dynamics simulations with PhysX than with ODE." I worked on the simulator at Anybots, and I'm pretty much convinced a different physics engine isn't going to magically make things better (in terms of using simulations to predict real robot behavior). The problem is more about getting every little physical detail of the actual robot -- and its interactions with the environment -- into the model. The particulars of how collision, joint constraints, and dynamics are numerically computed isn't what makes the problem really hard.

palish 18y ago
Yeah, after giving it some thought, it would be very tough to repurpose a realtime physics simulator like PhysX or ODE to be useful to Anybots. But that doesn't necessarily mean that it isn't practical to create a simulator suited for rapid prototyping of robots. (I'm just brainstorming ideas in the rest of this post; I'm not sure whether they'll work, but it's a fun problem to think about.) Perhaps it would be useful to write your own simulator that incorporates data about how the components of the robot behave in the real world, to get a more accurate simulation. By that I mean, the problem is that the robot model has so much complexity that it's difficult to simulate accurately. So instead of creating a mathematical model for how a compressed-air actuator behaves, for example, gather real-world data about how it actually behaves in the robot, then incorporate that data into the simulator. I mean, if the robot is walking slowly on a flat surface, then its individual components behave in the same ways each time the robot takes a step forward, right? Therefore it seems like the behavior of the whole robot can be reproduced and predicted in a simulator, since the behavior of each component is known (under controlled conditions, like when the robot is walking slowly). Brute force the problem by using large amounts of real-world data, is what I'm getting at.
- danbmil99 18y ago
  It's not that PhysX or ODE aren't good enough, it's that they only solve part of the problem. We did do some empirical analysis to get the pneumatic models. "Brute force the problem by using large amounts of real-world data, is what I'm getting at." Absolutely, that is Trevor's viewpoint, and I agree. Turns out to be somewhat difficult in practice. It's hard to even know what all the parameters should be, much less do all the work necessary to tune them. And all the while, the robot itself is changing. Joints get loose; servos get replaced. It's a constant battle. There is a feeling among hard-core robotics folks that this level of simulation is either impossible or not worth the trouble. I tend to disagree. I think we got pretty far with the Anybots simulator; we did manage to evolve a neural network that balances on the simulator, and also balances the real robot. It's definitely a 'hard' problem to get a simulator good enough to be usefully predictive.

palish 18y ago

Yeah, after giving it some thought, it would be very tough to repurpose a realtime physics simulator like PhysX or ODE to be useful to Anybots. But that doesn't necessarily mean that it isn't practical to create a simulator suited for rapid prototyping of robots. (I'm just brainstorming ideas in the rest of this post; I'm not sure whether they'll work, but it's a fun problem to think about.) Perhaps it would be useful to write your own simulator that incorporates data about how the components of the robot behave in the real world, to get a more accurate simulation. By that I mean, the problem is that the robot model has so much complexity that it's difficult to simulate accurately. So instead of creating a mathematical model for how a compressed-air actuator behaves, for example, gather real-world data about how it actually behaves in the robot, then incorporate that data into the simulator. I mean, if the robot is walking slowly on a flat surface, then its individual components behave in the same ways each time the robot takes a step forward, right? Therefore it seems like the behavior of the whole robot can be reproduced and predicted in a simulator, since the behavior of each component is known (under controlled conditions, like when the robot is walking slowly). Brute force the problem by using large amounts of real-world data, is what I'm getting at.

danbmil99 18y ago
It's not that PhysX or ODE aren't good enough, it's that they only solve part of the problem. We did do some empirical analysis to get the pneumatic models. "Brute force the problem by using large amounts of real-world data, is what I'm getting at." Absolutely, that is Trevor's viewpoint, and I agree. Turns out to be somewhat difficult in practice. It's hard to even know what all the parameters should be, much less do all the work necessary to tune them. And all the while, the robot itself is changing. Joints get loose; servos get replaced. It's a constant battle. There is a feeling among hard-core robotics folks that this level of simulation is either impossible or not worth the trouble. I tend to disagree. I think we got pretty far with the Anybots simulator; we did manage to evolve a neural network that balances on the simulator, and also balances the real robot. It's definitely a 'hard' problem to get a simulator good enough to be usefully predictive.