New RL method achieves 98.3% safe robot navigation with CVaR constraints
Post-training reachability verification catches tail risks that average costs miss.
Safe navigation for mobile robots under perception uncertainty remains a critical challenge. Existing safe reinforcement learning (RL) methods typically evaluate safety using average cumulative cost, which can mask dangerous tail-risk behaviors. Researchers from multiple institutions propose a framework that trains risk-sensitive policies through Conditional Value-at-Risk (CVaR) constrained optimization on an off-policy TD3 backbone. After training, they apply neural network reachability verification using Taylor Model analysis to compute action reachable sets under bounded observation uncertainty, yielding a safety rate metric that quantifies the proportion of evaluated states where the policy stays within safety margins.
In experiments across ten navigation scenarios and six baselines, the CVaR-constrained policies maintained larger safety margins from obstacles and achieved a 98.3% success rate, the highest safety verification rate among all compared methods. Critically, the authors found that average cost rankings and reachability-based safety rankings can diverge, showing that reachability verification captures risks missed by empirical cost metrics alone. The framework was also validated on a physical Clearpath Jackal robot, demonstrating successful sim-to-real transfer. The paper is available on arXiv (2605.14174).
- Uses CVaR (Conditional Value-at-Risk) constraints during training to focus on high-cost tail outcomes rather than average cost.
- Post-training reachability verification with Taylor Models computes safety margins under observation uncertainty.
- Achieved 98.3% success rate across 10 scenarios, outperforming 6 baselines; validated on a Clearpath Jackal robot.
Why It Matters
This method bridges the gap between training and formal safety guarantees, critical for deploying robots in cluttered real-world environments.