HAML: Humanoid Adversarial Multi-Skill Learning via a Single Policy

Translating large-scale motion datasets into robust, deployable humanoid controllers is a critical challenge in engineering informatics, primarily due to the scarcity of high-quality annotations, the risk of mode collapse in conditional generation, and the strict constraints of onboard computing hardware. This paper presents a deployable two-stage learning system that maps clip-level motion datasets to a single-policy multi-skill controller and its deployable counterpart. We adopt coarse one-hot skill labels that can be assigned automatically at the clip level with negligible manual effort, enabling scalable dataset construction. To prevent conditional discriminators from ignoring skill conditions, we inject mismatched (transition, label) pairs and introduce a condition-aware loss that explicitly penalizes incorrect transition–label associations, improving controllability and mitigating mode collapse. For real-world deployment, we further propose a two-stage training strategy: a privileged teacher policy is first trained in simulation and then distilled into a student policy that relies on stacked historical proprioceptive observations, ensuring robustness against sensing noise and latency without relying on external state estimation. Extensive evaluations in simulation and on real hardware demonstrate improved skill coverage, transition coverage, realism, and training efficiency across heterogeneous embodiments. With the onboard computer of a Unitree G1 robot, the distilled policy runs at 100 Hz with 15–25 ms latency, confirming the system’s engineering feasibility.