Event

LECO: Offline-Learned Direct Control Policy for Dynamic Target Tracking

1 reports · 1 sources

Bing Sun and other authors propose the LECO framework, which converts online trajectory optimization for dynamic target tracking into an offline-learned direct control policy, and uses a differentiable kinematics model to let trajectory-level objectives directly guide control learning. Under three simulated obstacle densities, LECO's root-mean-square linear jerk is 14.1%–31.2% lower than that of the comparison method with the best value on this metric, its tracking success rate is the highest, and its per-cycle computation time is 4.25–5.63 ms. At deployment, a predictive safety lock is included: when the predicted motion is unsafe or the target prediction is invalid, it switches to classical planning and control as a fallback. The authors state that experiments on two robot platforms further demonstrate tracking in indoor narrow corridors and outdoor forest environments. The above performance data come from simulation experiments; the safety lock and real-robot tracking are the design and experimental results reported by the authors.