Roxie
An open-source reinforcement learning framework written in JAX, focused on continuous control tasks in MuJoCo environments. It provides a unified hardware interface, seven built-in RL algorithms, and a fully compiled training loop.
Timeline (last 90 days)
Roxie is a reinforcement learning framework written entirely in JAX, designed specifically for continuous control tasks in MuJoCo environments.
Roxie was published on Reddit by developer Vittorio.
Roxie drives different hardware paths of MuJoCo through a single unified interface, so regardless of whether the underlying execution runs on GPU or pure CPU, developers face the same set of APIs.
Roxie has seven built-in continuous control algorithms: DDPG, TD3, D4PG, TD4, SAC, MPO, and PPO.
In Roxie, all agents share the same agent base class, converging algorithmic differences into minimal pluggable modules.
Roxie compiles both the acting and learning phases into a single XLA program, with no operations needing to return to the host side during runtime.
Roxie used a training budget of 500 million steps (500M-step) to test all 7 agents on 3 humanoid tasks in dm_control, covering both GPU and pure CPU execution units.
Roxie's pre-trained checkpoints have been uploaded to Hugging Face (vittorione/roxie-release-v1).
Roxie's framework code is hosted on GitHub (github.com/vittorione94/roxie).
Roxie focuses on the vertical domain of MuJoCo continuous control rather than being a large, all-encompassing general-purpose RL library.
All Facts (10)
Roxie focuses on the vertical domain of MuJoCo continuous control rather than being a large, all-encompassing general-purpose RL library.
50%UnverifiedRoxie's framework code is hosted on GitHub (github.com/vittorione94/roxie).
50%UnverifiedRoxie is a reinforcement learning framework written entirely in JAX, designed specifically for continuous control tasks in MuJoCo environments.
50%UnverifiedRoxie drives different hardware paths of MuJoCo through a single unified interface, so regardless of whether the underlying execution runs on GPU or pure CPU, developers face the same set of APIs.
50%UnverifiedRoxie has seven built-in continuous control algorithms: DDPG, TD3, D4PG, TD4, SAC, MPO, and PPO.
50%UnverifiedIn Roxie, all agents share the same agent base class, converging algorithmic differences into minimal pluggable modules.
50%UnverifiedRoxie compiles both the acting and learning phases into a single XLA program, with no operations needing to return to the host side during runtime.
50%UnverifiedRoxie used a training budget of 500 million steps (500M-step) to test all 7 agents on 3 humanoid tasks in dm_control, covering both GPU and pure CPU execution units.
50%UnverifiedRoxie was published on Reddit by developer Vittorio.
50%UnverifiedRoxie's pre-trained checkpoints have been uploaded to Hugging Face (vittorione/roxie-release-v1).
50%