Hangzhou-based DEEP Robotics has open-sourced a full IsaacLab-based reinforcement-learning training and deployment library for its quadruped and humanoid platforms, giving academic and industrial teams a ready-to-run blueprint that goes all the way from simulation to real hardware.
Three official envs, one repo
The rl_training repository, published on GitHub in early September, exposes three official environments: Rough-Deeprobotics-Lite3-v0 for the Lite3 quadruped, Rough-Deeprobotics-M20-v0 for the M20 (branded as LYNX M20 for overseas markets) and Amp-Flat-Deeprobotics-DR02-v0 for the DR02 platform. Each env ships with configs, PPO/AMP training scripts (via RSL-RL), simulation assets, multi-GPU acceleration hooks and one-command ONNX export so operators can deploy the trained policies to MuJoCo or the real robot without an Isaac Sim runtime.
Why open-sourcing the loop matters
Modern legged-robot policies live and die by the sim-to-real pipeline: reward shaping, domain randomisation, actuator models and the sim-to-hardware bridge are exactly the pieces academic teams spend months rebuilding for every new platform. DEEP Robotics is short-circuiting that by publishing a reference stack anchored on NVIDIA's IsaacLab (with RSL-RL 5.0.1 and Isaac Sim 5.1), plus tutorial videos on Bilibili and YouTube.
A hardware-plus-open-source pincer
The release lands weeks after DEEP Robotics filmed its DR02 tackling outdoor stairs in a single take and as it moves toward a STAR Market IPO, and it plugs into a broader open-humanoid push that includes Robocurve's GPT-6 Astra benchmarks and Caterpillar–FieldAI's NVIDIA Omniverse work.
Reporting based on the DeepRoboticsLab/rl_training GitHub repository and DEEP Robotics tutorial materials.