Delay-Resilient RL Framework for Stable Robot Teleoperation
A study published on arXiv (2605.15480) presents a novel hybrid control strategy known as delay-resilient RL, aimed at mitigating stochastic communication delays in teleoperation. Such delays lead to interruptions in signals that compromise control stability, and traditional reinforcement learning techniques often falter with delayed inputs, resulting in excessive chattering. The introduced framework combines a state estimator based on Long Short-Term Memory (LSTM) with a residual RL policy. This LSTM generates smooth, continuous state estimates from delayed data, allowing the RL agent to develop a torque compensation policy that optimizes both tracking precision and velocity smoothness. Tests conducted on Franka Panda robots demonstrate that this method significantly surpasses leading baselines, providing stable and reliable teleoperation even in high-delay scenarios.
Key facts
- arXiv paper 2605.15480 proposes delay-resilient RL for teleoperation
- Stochastic delays cause signal discontinuities and chattering
- Framework uses LSTM for state estimation and residual RL policy
- LSTM reconstructs smooth state estimates from delayed observations
- Residual policy balances tracking accuracy and velocity smoothness
- Tested on Franka Panda robots
- Outperforms state-of-the-art baselines
- Ensures robust teleoperation under high delays
Entities
Institutions
- arXiv