Athena-WBC: Capability-Aligned Policy Experts for Long-Tail Humanoid Whole-Body Control

2026-07-06

AI Summary

Existing whole-body control methods for large-scale humanoid robots often fail on long-tail motions involving high-dynamic transitions and critical balancing due to a mismatch between policy capabilities and motion demands. This work proposes Athena-WBC, a capability-aligned teacher–student framework that employs two specialized experts—one optimized for trajectory tracking under dynamic conditions and the other for training stability during balance-intensive tasks. The approach integrates routing-based distillation, DAgger imitation learning, constraint-aware objectives (without conservative cost shaping), gravity-based curricula, and reinforcement learning fine-tuning to distill these expert policies into a single deployable controller. Athena-WBC significantly enhances recovery robustness on long-tail motions and improves tracking performance on unseen actions, outperforming the strong baseline SONIC while requiring only a minimal set of expert policies.

Abstract

Large-scale humanoid motion-tracking controllers are commonly improved by reallocating training effort: difficult motions are sampled more often, isolated into smaller subsets, or assigned to specialized experts. We show that this view is incomplete. In strong whole-body-control baselines, a residual set of feasible training clips remains unsolved even under targeted training, especially for high-dynamic transitions and balance-critical motions. These failures arise not only from insufficient exposure, but from a mismatch between the motion demands and the effective capability induced by the default training recipe. We propose Athena-WBC, a compact teacher-student pipeline with capability-aligned policy experts for long-tail humanoid whole-body control. Dynamic experts use a tracking-focused, constraint-aware objective that removes conservative effort and temporal-control penalties while preserving physical feasibility constraints; balance experts use a gravity curriculum to improve early-training survivability. The resulting privileged teachers are motion-routed for DAgger distillation and then compressed into a single controller with deployable observations followed by RL fine-tuning. Experiments on a full-size humanoid show improved recovery of training-set long-tail motions and better held-out tracking than a strong SONIC-recipe baseline, using only a small number of experts.

Problem

• humanoid whole-body control
• long-tail motions
• capability mismatch
• high-dynamic transitions
• balance-critical motions

Innovation

• capability-aligned experts
• whole-body control
• teacher-student distillation
• long-tail motion
• constraint-aware policy