A LITTLE REINFORCEMENT.
Intelligence,
in motion.
Three shots. A window into my work.
Yifan Su 苏逸凡
Teaching robots to learn,
compete, and move.
Member of Technical Staff at Skild AI in San Francisco. Incoming Robotics Ph.D. student at Carnegie Mellon, advised by Deepak Pathak and Abhinav Gupta.
I work on reinforcement learning for control and manipulation, sim-to-real transfer, and planning and coordination for multi-agent systems.
Education & research background
I received my B.S. in Computer Science from CMU in May 2025. At CMU, I worked with Jiaoyang Li on robust multi-agent path execution, with Deepak Pathak on scaling reinforcement learning for manipulation and locomotion, and with Sebastian Scherer and Micah Corah on multi-drone reconstruction at AirLab.
Learning
through self-play.
I independently developed Skild AI’s humanoid soccer demo, training policies through self-play in NVIDIA Isaac Sim and transferring them to real hardware.
With a scoring objective and evolving opponents, dribbling, shielding, and tackling emerged without handcrafted rewards for these behaviors. After 140 years of simulated play, we deployed the policy on a real humanoid.
Explore Physical Self-Play ↗Beyond
the sandbox.
Agile whole-body control on real hardware: a humanoid robot traversing obstacles, jumping onto boxes, and moving through the world.
I lead Unitree G1 sim-to-real deployment, including simulation fidelity, domain randomization, and control tuning, and built the parkour deployment pipeline and safety filters for latency, sensor noise, and hardware constraints.
Watch Parkour ↗THE LATEST
Along the way.
- NEXT
Joining CMU Robotics as a Ph.D. student with Deepak Pathak and Abhinav Gupta.
- AAAI 2026
BTPG-max accepted to AAAI 2026.
- MAY 2025
Graduated from Carnegie Mellon; joined Skild AI as a Member of Technical Staff.
- MAR 2025
Evolutionary Policy Optimization released on arXiv.
- FEB 2024
Presented our BTPG paper as an oral at AAAI in Vancouver.