Repository navigation
PACE v0.2.0
⚠️ Breaking: actuator delay now matches the paper
The actuator delay was applied to the computed torque, while the paper describes a delay of the commands. In v0.2.0 the delay acts on the position/velocity targets and feed-forward torques (delay_mode="command", the new default). The PD loop runs on current measurements, which matches drives with their own low-level control loop and stays stable for large delays.
Migration: the results in the paper were produced with the previous implementation. To reproduce them, or to keep using parameters identified with v0.1.x, set:
PaceDCMotorCfg(..., delay_mode="effort")Delays identified with v0.1.x are not directly transferable to the new default, so please re-run the identification.
✨ New: separate static and dynamic friction
Set separate_dynamic_friction=True in your PaceCfg to identify break-away (static) and dynamic friction separately. Dynamic friction is identified as a ratio of the static friction (friction_ratio_bounds, default [0, 1]), which adds one parameter per joint and guarantees dynamic ≤ static friction. With the flag off (default), friction stays pure Coulomb as before. Requires Isaac Sim ≥ 5.0.
The ANYmal example now uses the split (61 instead of 49 parameters), and its synthetic data uses a dynamic friction of half the static friction.
📄 Paper published in IJRR
Towards Bridging the Gap: Systematic Sim-to-Real Transfer for Diverse Legged Robots is now published in The International Journal of Robotics Research (doi:10.1177/02783649261459628, open access on arXiv). The README and docs contain the updated citation.
Other changes
- Fixed plot_trajectory.py printing the wrong entry as delay
- Documentation updated for both features, including the parameter layout and deployment steps
- GitHub star history page in the docs
- License split into license and copyright notice
Contributors
Thanks to @olivefi for the delay fix (#25)!
Full changelog: v0.1.2...v0.2.0