JoltGym

A high-performance, MuJoCo-compatible physics simulation engine for reinforcement learning, built on Jolt Physics with Vulkan rendering and Python (Gymnasium) bindings.

Key Features

• Jolt Physics backend -- C++17, deterministic, excellent multicore scaling
• MJCF compatibility -- loads MuJoCo XML models (HalfCheetah, Humanoid)
• Gymnasium API -- drop-in env.reset() / env.step() / env.render() interface
• Vulkan renderer -- SDL2 window + Dear ImGui debug overlay (macOS via MoltenVK, Linux, Windows)
• Offscreen rendering -- headless rgb_array mode for training servers
• WorldPool -- N parallel PhysicsSystem instances stepped from C++, ~73K env-steps/sec
• Multi-agent -- multiple robots in a single shared world with collision
• Zero-copy Python -- pybind11 + NumPy, hot path entirely in C++

Architecture Overview

Python (Gymnasium API)
  +-- pybind11 bindings (zero-copy NumPy)
       +-- C++ Core
            |-- MJCF Parser (tinyxml2)
            |-- Physics (Jolt Physics)
            |-- Renderer (Vulkan + SDL2 + ImGui)
            +-- WorldPool (N parallel PhysicsSystem instances)

Environments

Environment	Description	Obs Dim	Act Dim
JoltGym/HalfCheetah-v0	2D planar cheetah locomotion	17	6
JoltGym/Humanoid-v0	3D bipedal humanoid locomotion	45	17
JoltGym/CheetahRace-v0	N multi-agent cheetahs racing	N*17	N*6

Quick Example

import joltgym

env = joltgym.make("JoltGym/HalfCheetah-v0")
obs, info = env.reset(seed=42)

for _ in range(1000):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        obs, info = env.reset()

env.close()

Performance

JoltGym scales to 3.8x faster than MuJoCo at 64+ parallel environments via C++ parallel stepping, eliminating Python subprocess overhead entirely.

Environments	JoltGym (C++ WorldPool)	MuJoCo (AsyncVectorEnv)
1	11,935 sps	17,477 sps
8	33,292 sps	11,024 sps
64	64,503 sps	18,608 sps
256	73,606 sps	18,287 sps

See Performance for full benchmarks and analysis.

License

MIT