New hybrid RL method stabilizes robot teleoperation under delays

Stochastic communication delays in teleoperation cause signal discontinuities and high-frequency chattering, undermining control stability. Conventional reinforcement learning methods falter because delayed observations mess up the state-action mapping. To solve this, researchers Kaize Deng and Zewen Yang developed a hybrid control framework called delay-resilient RL. It integrates a Long Short-Term Memory (LSTM) network as a state estimator to reconstruct smooth, continuous state estimates from the delayed and jittery sensor readings. On top of that, a residual RL policy learns to add just enough torque compensation to balance tracking accuracy with velocity smoothness, without requiring retraining from scratch for every delay profile.

Experimental validation on real Franka Panda robots showed the approach significantly outperforms state-of-the-art baselines (like pure RL or model-predictive control) across multiple stochastic delay distributions. The system maintained stable, chatter-free teleoperation even under high-variance delays that would normally cause oscillations. This work, accepted at the 23rd IFAC World Congress 2026, points toward more practical teleoperation systems for remote surgery, hazardous environment handling, and space robotics where network latency is unpredictable.