Adaptive Smooth Tchebycheff framework helps robots master conflicting objectives
Researchers solve a decade-old challenge in multi-objective robot learning with adaptive optimization.
Researchers Murillo-Gonzalez, Ali, and Liu propose Adaptive Smooth Tchebycheff Attention for multi-objective policy optimization. Their framework dynamically modulates the optimization landscape using a conflict-driven controller that adjusts smoothness based on real-time gradient interference. Tested on a robotic stealth visual search task (a proxy for monitoring protected/fragile ecosystems), it robustly discovers Pareto-optimal policies in non-convex regions inaccessible to linear baselines and unstable for static non-linear methods. The work will appear at RSS 2026.
- Linear scalarization cannot recover solutions in non-convex Pareto front regions, a known limitation in multi-objective RL.
- The framework uses a conflict-driven controller to dynamically adjust optimization smoothness based on gradient interference.
- Validated on a robotic stealth visual search task for ecosystem monitoring, outperforming both linear and static non-linear baselines.
Why It Matters
Enables robots to make smarter trade-offs between competing goals, critical for autonomous systems in complex environments.