Safety Filtering for High-Dimensional Stochastic Systems via Particle Reachability Value Functions
Abstract
We are developing a safety control and filtering method for high-dimensional stochastic systems using reachability analysis, control barrier functions, and machine learning. If you are interested in this work, please contact Hyeontae Sung.
Preliminary Results
Our method achieves better safety than the chance-constrained MPC in social navigation tasks.
Chance-constrained MPC vs. our method in a social navigation scenario with one human obstacle. Our method avoids the human obstacle with a larger distance.
Table 1: Social navigation results.
| Method | n = 1 | n = 5 | n = 10 | |||
|---|---|---|---|---|---|---|
| Timeout (%) | Collision (%) | Timeout (%) | Collision (%) | Timeout (%) | Collision (%) | |
| CC-MPC | 2 | 13 | 7 | 23 | 11 | 19 |
| Ours | 1 | 1 | 11 | 3 | 19 | 10 |
CC-MPC denotes the chance-constrained MPC. n = 1, 5, 10 indicates the number of human obstacles. Timeout denotes the percentage of times that the method failed to reach the goal within the time limit. Collision denotes the percentage of times that the method caused a collision.