Reusing Locomotion Training Recipes Across Humanoids via Body-Aligned Virtual Joints

TL;DR A kinematic layer lets humanoid locomotion policies act in a shared body-aligned joint space, so one unmodified RL recipe trains well on three humanoids with different leg-axis layouts. On a Unitree G1 it cut heading drift from 5.31 to 0.89 deg/s of yaw rate.
What they do
Humanoids mount their leg joint axes differently (oblique hips, inclined pitch axes, different orderings). RL locomotion policies output native joint angles, and action scales, bounds, gains and penalties are all set per native joint, so a recipe tuned on one robot does not carry over. The authors let the policy act and observe in a common body-aligned joint space (hip yaw-roll-pitch, knee, ankle roll-pitch), leaving each robot’s transmission and low-level control unchanged.
How it works
- Virtual joints from link frames. Virtual hip and ankle angles come from the relative orientation of the bounding links (pelvis-thigh, shank-foot), calibrated to zero at the nominal pose and derived from the URDF.
- Mapping to native targets. Virtual targets become native ones by orientation matching; a coordinate Jacobian maps joint velocities and converts shared virtual-space gains into native PD gains.
- Post-hoc insertion. The policy trains on a copy of the robot whose leg axes are replaced by the canonical ones, and the map is applied only at deployment. Applying it inside the training loop was unstable.
- One recipe, adapted from Unitree’s G1 velocity-tracking setup, is trained in Isaac Lab for the G1, PND Adam and a custom humanoid, and tested in MuJoCo.
Results
- On Adam and the custom robot it had the lowest velocity error (0.085 and 0.029 m/s) and lateral drift; on the G1, whose hip is nearly body-aligned, all interfaces tracked similarly.
- Mean toe-out on Adam was 0.5°, against 4.8° for native joints and 14.2° for a task-space interface.
- Training was unstable for at most 0.1% of iterations, against 15.3% for the in-loop variant on Adam.
- G1 hardware, 6 straight-walking rollouts per interface: yaw rate 0.89 ± 0.54°/s against 5.31 ± 1.90°/s for native joints. Every proposed rollout reached the end of the walkway (13.8 ± 1.1 m); native rollouts drifted off after 5.1 ± 1.7 m. The map costs about 0.2 ms per step on a Jetson Orin NX.
Limitations
Legs-only velocity tracking, three seeds per configuration, and hardware tests on one robot, with drift measured as IMU yaw rate because base position was not logged. The authors note that explaining the hardware gap would need further hardware ablations.
Figures



Abstract
Reinforcement learning has enabled robust humanoid locomotion, yet training recipes (rewards, architectures, hyperparameters) remain hard to reuse across robots with different native joint coordinates, since policies output native joint angles while action scales, bounds, gains, and penalties are defined in those robot-specific coordinates. Existing transmission models convert native targets into actuator commands but do not standardize the coordinate interface exposed to the policy. We introduce a body-aligned virtual joint space upstream of this mapping: policy actions and observations follow a common anatomical convention, and a deterministic kinematic layer maps virtual targets to native ones, leaving the existing transmission and low-level control stack unchanged. Under a single unmodified recipe on three humanoids with different leg-axis layouts, it is the only interface that tracks accurately, keeps the feet aligned with the travel direction, and trains stably on all three, and on Unitree G1 hardware it reduces heading drift by nearly a factor of six.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
