Keeping Jerk in Check: Jointly Trained Perception-Gated Experts for Smooth Terrain Transitions
Keeping Jerk in Check: Jointly Trained Perception-Gated Experts for Smooth Terrain Transitions
This research presents CTS-MoE, a multi-task reinforcement learning framework that jointly trains a perception-gated dense mixture-of-experts actor with task-specific critics to enable smooth transitions across discontinuous terrain. The approach uses soft routing to compose expert knowledge continuously and transfers composition from privileged terrain information to onboard depth for deployment without labels. Evaluated on a Unitree Go1, the policy achieves 93.3% connected-course success compared to 73.3% and 53.3% for hierarchical and DAgger-distilled baselines, reduces mean action jerk by approximately one-third, and completes an extended outdoor course in all 10 trials across four terrain types.
The paper introduces CTS-MoE, a jointly trained perception-gated mixture-of-experts policy that improves connected-course success to 93.3% and reduces action jerk on a Unitree Go1 across four outdoor terrain types.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
trials completed
10
Reported trials
View original evidence
ourse success to 93.3% from 73.3% and 53.3%, lowers mean action jerk by about a third across the course, and avoids the switching transient that hard selection introduces at boundaries. On a Unitree Go1, it completes an extended outdoor course in 10/10 trials and transfers to four outdoor terrain types. Project Website: https://cts-moe.github.io/ .
Open source S5trials total
10
Reported trials
View original evidence
ourse success to 93.3% from 73.3% and 53.3%, lowers mean action jerk by about a third across the course, and avoids the switching transient that hard selection introduces at boundaries. On a Unitree Go1, it completes an extended outdoor course in 10/10 trials and transfers to four outdoor terrain types. Project Website: https://cts-moe.github.io/ .
Open source S5terrain types transferred
4
View original evidence
ourse success to 93.3% from 73.3% and 53.3%, lowers mean action jerk by about a third across the course, and avoids the switching transient that hard selection introduces at boundaries. On a Unitree Go1, it completes an extended outdoor course in 10/10 trials and transfers to four outdoor terrain types. Project Website: https://cts-moe.github.io/ .
Open source S5
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
d critics isolate heterogeneous objectives so that specialization survives joint optimization. A concurrent teacher-student formulation transfers this composition from privileged terrain information to onboard depth, so that at deployment routing is inferred from perception alone, with no terrain label, selector, or distillation stage. Against hierarchical and DAgger-distilled baselines, the policy raises connected-c
Open source S4
ourse success to 93.3% from 73.3% and 53.3%, lowers mean action jerk by about a third across the course, and avoids the switching transient that hard selection introduces at boundaries. On a Unitree Go1, it completes an extended outdoor course in 10/10 trials and transfers to four outdoor terrain types. Project Website: https://cts-moe.github.io/ .
Open source S5
Keeping Jerk in Check: Jointly Trained Perception-Gated Experts for Smooth Terrain Transitions
Open source S1
Source：arXiv Robotics — research abstracts · arxiv.org