RSL-RL is the underlying reinforcement learning engine that powers much of the modern legged locomotion research conducted at ETH Zurich and beyond. While many general-purpose RL libraries exist, RSL-RL was purpose-built for the unique demands of simulating and controlling articulated legged robots at scale. It implements a streamlined variant of Proximal Policy Optimization (PPO) that is optimized for GPU batch processing, enabling training runs with thousands of parallel agents without the CPU bottlenecks common in standard implementations. The design philosophy behind RSL-RL is minimalism combined with performance. The entire training loop is implemented in pure PyTorch, with tensor operations carefully shaped to maximize GPU utilization. Unlike heavier libraries that abstract every policy gradient variant into a generic interface, RSL-RL exposes a clean, narrow API focused on actor-critic methods with Generalized Advantage Estimation (GAE). This deliberate simplicity makes it easy to inspect, debug, and extend—an important consideration when working with physical robots where a subtle bug in the advantage computation can lead to catastrophic falls in the real world. RSL-RL integrates seamlessly with NVIDIA Isaac Gym, the GPU-accelerated physics simulator that allows whole armies of robots to be simulated in parallel. The library handles the coordination between Isaac Gym’s batched environment API and the PPO update loop, taking care of rollout collection, advantage computation, and multi-epoch policy updates under a single configuration file. It also supports features particularly important for locomotion: asymmetric actor-critic training (where the critic receives privileged state information unavailable to the actor), observation normalization, and automatic entropy scheduling to encourage policy exploration without manual tuning. The codebase has become the default RL backend for Legged Gym and is used in numerous follow-up projects spanning quadruped locomotion, humanoid balancing, and dexterous manipulation. Its lightweight footprint and high throughput have made it a go-to choice for researchers who need reliable, fast PPO training without the overhead of larger frameworks like RLlib or Stable Baselines3. For anyone entering the legged robot learning space, familiarity with RSL-RL is effectively a prerequisite.