K-Scale RL Hackathon 0.5 Pilot: Train & Deploy on Real Robots
Overview
Have you always wanted to program and train reinforcement learning (RL) policies and deploy them onto real humanoid robots? Join us this weekend in-person to try K-Sim, our high-performance, open-source RL training library, and deploy your policies directly onto our robots.
We accept submission in-person and online. Deployment results will be live streamed. Only limited number of people will be invited to be in person. Please submit any issues on GitHub.
For any questions please directly message JX on Discord at: jingxiangmo
Introducing K-Sim & KOS
K-Sim is an open-source, high-performance reinforcement learning framework optimized for training humanoid robot locomotion, manipulation, and real-world deployment. Built with MJX, K-Sim simplifies training, evaluation, and sim-to-real deployment onto actual robots. With easy setup (pip install ksim) and compact, single-file scripts for PPO and AMP, users can quickly train policies and deploy them to real robots in under an hour.
Key Features:
End-to-End Pipeline: Deploy to robots with a single command via KOS.
Differentiable GPU Physics: High-performance simulations using JAX and Mujoco XLA.
Single-file Scripts: Simplify experimentation and debugging.
Advanced Architectures: Built-in support for stateful policies (e.g., RNNs, Transformers).
Interactive Visualizer: Real-time policy feedback and refinement.
KOS-Sim is a digital twin and model evaluator for the K-Scale Operating System (KOS), using the same gRPC interface as the real robot. Easily test and refine your models in simulation, then deploy seamlessly by simply changing the client IP address.
Why Join?
Excited about programming humanoid robots and deploying RL policies onto real hardware? Join us to:
Train and deploy your RL policies for real-world tasks such as walking, dancing, and performing complex tricks.
Compete and showcase your skills on our open-source leaderboard to advance humanoid robot capabilities: leaderboard.kscale.dev
Experiment easily with provided Google Colab notebooks or use your own GPU resources.
Collaborate with leading roboticists and machine learning researchers.
About
We build open-source humanoid robots in Palo Alto. https://www.kscale.dev/
We own our stack from shipping state of the art machine learning models trained on our infrastructure, building the operating system, designing the hardware, and manufacturing it.
K-Sim GitHub: https://github.com/kscalelabs/ksim
Humanoid Benchmark Repo: https://github.com/kscalelabs/kscale-humanoid-benchmark
Leaderboard & Submission: https://leaderboard.kscale.dev/
Join now to innovate and accelerate the future of humanoid robot locomotion!