Yehonathan Litman

🚀 New open-source project! I’ve extended the learning framework from my previous works, GfR (RSS 2026) and HIL (TOG 2026), to two representative tasks on the Unitree G1: box moving and box climbing, built on top of Holosoma. The same learning framework applies to both locomotion and loco-manipulation tasks. Using only a single reference motion clip for each task, the learned policies generalize across a wide range of task conditions and goals. The key idea (from GfR and HIL) is simple: 1️⃣ We found that a policy can achieve high-quality motion tracking without receiving the target reference motion as actor input. Instead, task goals are extracted from the reference motion and provided to the actor together with the robot state. The reference itself is used to define the tracking reward. The policy is conditioned on task goals rather than the reference motion, and the learned behavior can easily transfer to new task conditions. 2️⃣ Based on this, we develop a hybrid multi-task training setup: • In the tracking task, goals are extracted from the reference motion and the policy is trained with tracking rewards. • In the general RL task, goals and task conditions are randomly sampled, and the policy is trained with sparse goal-reaching rewards. • The actor always receives the same observation structure across tasks, while the critic is provided with task-specific signals. The same policy learns from both tracking and general RL tasks, transferring behaviors acquired through motion tracking to more general goal-driven settings. This project provides a practical example of how the GfR/HIL methodology can be adapted to new tasks. Note: this is an unofficial research implementation and extension, not the official code release for either GfR or HIL. 🔗 github.com/jiashunwang/Hybri…