RL, Sim-to-Real & Locomotion
Reinforcement learning (RL) is trial-and-error training of a robot controller, usually in simulation before transfer to hardware. In 2021 an ANYmal quadruped learned to walk on flat ground in under four minutes on one GPU Rudin et al. 2021Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning. Nikita Rudin, David Hoeller, Philipp Reist et al., CoRL 2021..
- RL for Robotics
Sample efficiency decides which RL algorithm a robot can use.
- Why RL Won Locomotion but Not Manipulation
Walking is cheap to simulate; contact-rich manipulation is not.
- Massively Parallel Sim RL
GPU-parallel simulation with Isaac Lab, Newton, MJX and Brax.
- Sim-to-Real Transfer
Domain randomization, distillation, system identification and real-to-sim reconstruction.
- Legged Locomotion Lineage
How learned gaits became the default for legged robots.
- Humanoid Whole-Body Control
How humanoids track retargeted human motion.
- Reward Design and the MPC Debate
Hand-tuned and language-model rewards, and the MPC debate.
- Offline Reinforcement Learning for Robotics
CQL and TD3+BC, and when behavior cloning is the stronger baseline.