Which physics engines are supported?

Ksim supports both the standard MuJoCo engine (CPU-based) and the MJX engine (JAX-native for GPU/TPU performance).

How are rewards defined in ksim?

Rewards are defined through a modular system using composable components, often utilizing exponential kernels and curriculum-based scaling to guide the learning process.

What is ksim used for?

Ksim is a reinforcement learning training library specifically designed for developing control policies for humanoid robots, focusing on locomotion and manipulation tasks.

Can ksim help with Sim2Real transfer?

Absolutely. It includes comprehensive modules for domain randomization, noise injection, and action latency modeling to ensure policies are robust enough for physical hardware.

Does ksim support GPU acceleration?

Yes, ksim utilizes MJX (MuJoCo in JAX) to provide JAX-native GPU and TPU acceleration, allowing for massive parallelization of physics simulations.

K-Scale Ksim Robotics RL

Name: K-Scale Ksim Robotics RL
Author: plurigrid

byplurigrid

•

Data Science & ML

Trains humanoid robot locomotion and manipulation policies using a JAX-native Reinforcement Learning framework built on MuJoCo.

K-Scale ksim is a specialized modular framework designed for high-performance reinforcement learning in robotics. It leverages MuJoCo and MJX (JAX-native MuJoCo) to provide hardware-accelerated physics simulations, enabling rapid training of walking and manipulation policies for humanoid robots. With built-in support for domain randomization, curriculum learning, and stateful observation systems, ksim facilitates seamless Sim2Real transfers and complex control loop implementations through a type-safe, JIT-compiled architecture.

Key Features

01Modular reward systems with composable, curriculum-scaled components.

02Domain randomization and disturbance injection for robust Sim2Real transfer.

03Advanced action processing including latency modeling and motor dynamics.

048 GitHub stars

05JAX-native physics simulation with MJX for high-speed GPU-accelerated training.

06Integrated curriculum learning schedules for managing training progression.

Use Cases

01Optimizing control policies for hardware deployment using Sim2Real techniques.

02Developing and training walking policies for bipedal humanoid robots.

03Simulating complex robotic manipulation tasks with realistic physics and sensor noise.

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add plurigrid/asi kscale-ksim

For use in Claude.ai and ChatGPT

Download Skill