Reinforcement Learning for Low-Cost Quadrupedal Robots with Delayed Feedback
A new paper on arXiv (2607.26434) addresses the sim-to-real gap in deploying reinforcement learning on cost-constrained quadrupedal hardware. The authors identify transport latencies and noisy motor feedback as key challenges, particularly on platforms like the Mini Pupper 2, where a measured >50 ms transport delay transforms the locomotion task into a partially observable Markov decision process. They propose a biologically inspired approach using a forward model of average actuator delay paired with a time-aware neural network, which learns a central pattern generator (CPG) to achieve robust locomotion. The work expands reinforcement learning capabilities on low-cost robotic platforms.
Key facts
- arXiv paper 2607.26434 addresses sim-to-real gap for low-cost quadrupedal robots
- Mini Pupper 2 platform has >50 ms transport delay
- Delay transforms locomotion into partially observable Markov decision process
- Biologically inspired approach uses forward model of actuator delay
- Time-aware neural network learns central pattern generator (CPG)
- Method achieves robust locomotion on cost-constrained hardware
- Focus on closing sim-to-real gap for reinforcement learning
- Published on arXiv with cross announcement type
Entities
Institutions
- arXiv