search

utiasDSL / safe-control-gym

PyBullet CartPole and Quadrotor environments—with CasADi symbolic a priori dynamics—for learning-based control and RL
https://www.dynsyslab.org/safe-robot-learning/
MIT License
560 stars 123 forks source link

Please explain the constraints #105

Closed MukilSaravanan closed 1 year ago

MukilSaravanan commented 1 year ago

Please explain the constraints that need to be satisfied along with maximizing the rewards. What does info['constraint values'] mean?

JacopoPan commented 1 year ago

Hello @MukilSaravanan

are is this a question relating the IROS 2022 competition?

The constraints are specified in the YAML files, e.g.

https://github.com/utiasDSL/safe-control-gym/blob/a0e7fa2b14a822f62d97a5cc59a6198317b6645a/competition/level3.yaml#L128-L162

simply means that there are upper and lower bound for the input (identical to the action_space) and the x, y, z position of the Crazyflies (bounding it to be within a flyable arena of 6 by 6 by 2.1 meters)

JacopoPan commented 1 year ago

info['constraint values'] are the numerical evaluations of those constraints (meaning that any negative value in that entry of the dictionary indicates a violation.

JacopoPan commented 1 year ago

I'm also tagging @adamhall if you have further implementation-specific questions.

MukilSaravanan commented 1 year ago

info['constraint values'] are the numerical evaluations of those constraints (meaning that any negative value in that entry of the dictionary indicates a violation.

It's actually the other way. ie, if any of those numerical evaluations of constraints is positive, it is said to be violating

JacopoPan commented 1 year ago

info['constraint values'] are the numerical evaluations of those constraints (meaning that any negative value in that entry of the dictionary indicates a violation.

It's actually the other way. ie, if any of those numerical evaluations of constraints is positive, it is said to be violating

I'm sure you're right (@adamhall can confirm or refute), I don't remember the sign convention by heart. The single boolean in info[constraint_violation] is intended to avoid any ambiguity.