The humanoid RL challenge
Train an RL policy tonight; watch it run on a real humanoid tomorrow.
Start RL training and zero-shot sim-to-real transfer now with K-Sim
K-Sim is an open-source library for GPU-accelerated robot learning and sim-to-real transfer, made for RL whole-body control from simple walking to complex human imitation.
Get started in <5 minutes
Got stuck along the way? Ask any questions in our Discord.
Join the challenge
We're building a leaderboard for anyone interested in rapidly moving from programming and training humanoid robots in simulation to seeing their ideas on real machines, the next day.
All submissions that pass our sim-to-sim evaluation will earn a spot on the leaderboard. We deeply appreciate your contributions in helping us advance our mission.
More in-depth evaluation criteria will be released soon.
Our competition will feature exciting prizes—ranging from fun company-branded merch to free access to the full-size K-Bot robot! Stay tuned for more details.
Challenges
We're planning to announce bi-weekly challenges in the future. Every week, we will deploy top policies on the real robot, which we'll livestream.
Basic Walk
Train an omnidirectional walking policy with velocity > 1m/s
Uneven Terrain
Survive walking across 100m Perlin hills and stairs
Push Recovery
Keep torso upright after random shoves of at least 50 Newtons
Human Motion Imitation
Track a 30-sec motion capture clip of human dancing and walking
Leaderboard coming soon. Star our repo for updates.
Go to GithubHow can I submit?
Export your policy to a K-Infer model
Evaluate in sim2sim and upload a Youtube video
Upload your submission to our Google Form
Post the link in #benchmark-submission on Discord