Hierarchical Gating Action Chunking Transformer for Joint Waist-Arm Humanoid Manipulation

Abstract

Humanoid manipulation requires coordinating global posture and fine motor control along a serial kinematic chain. However, most upper-body imitation learning policies focus on the arms and hands while treating the waist as an auxiliary component, leading to amplified end-effector errors and limited workspace coverage. We propose the Hierarchical Gating Action Chunking Transformer (HIG-ACT), a structure-aware extension of ACT that explicitly incorporates waist dynamics. HIG-ACT employs a hierarchical decoding architecture and a query-conditioned context gating mechanism to selectively regulate the influence of global waist posture on arm-hand prediction. Using whole-body demonstration data collected via an Apple Vision Pro-based teleoperation system, we evaluate HIG-ACT on waist-dependent manipulation tasks. The proposed method consistently outperforms ACT, and ablation studies confirm the effectiveness of both hierarchical decoding and gating in improving stability and success rates. © 2026 IEEE.