SymACT: a fixed symmetry consistency constraint method for bimanual robotic manipulation

Purpose This study aims to investigate whether fixed left-right symmetry can serve as an effective structural prior for Action Chunking Transformer (ACT)-based bimanual imitation learning and examines the task conditions under which such a prior improves or degrades policy performance. Design/methodology/approach SymACT applies predefined left-right mirror transformations to camera views, proprioceptive states and actions and trains a shared ACT policy with an action-level consistency loss. Four RoboTwin 2.0 tasks are evaluated using three training seeds and 100 fixed evaluation scenes per seed. The study also includes symmetry-weight and mirror-only ablations, end-effector trajectory analysis and three SO-ARM101 tasks with 50 physical trials per method and task. Findings In simulation, SymACT changes mean success rates by −24.0 percentage points on Handover Block and by + 6.0, +12.3, and + 6.0 percentage points on Lift Pot, Handover Mic and Grab Roller, respectively. On the physical platform, observed success rates change from 74% to 86% on Lift Box, 62% to 48% on Place Block and 70% to 72% on Handover Object. The results show task-dependent effects, with both improvements and degradation under a fixed symmetry consistency constraint. Originality/value The study provides an ACT-specific implementation of fixed mirror consistency and a multi-seed, multi-task characterization of where this structural prior helps or harms bimanual manipulation, bimanual robotic manipulation, ACT, fixed symmetry consistency, structural prior.