VAE-MPPI: Closed-Chain Constrained Trajectory Optimization for Dual-Arm Manipulation

VAE-MPPI: 양팔 로봇 조작을 위한 폐쇄 연쇄 제약 궤적 최적화

Coordinated dual-arm manipulation, where two robots jointly interact with a shared object, forms a closed-chain system that imposes strict kinematic equality constraints on the robot configuration. Conventional Model Predictive Path Integral (MPPI) control addresses such constraints only through soft penalty terms, making strict feasibility guarantees fundamentally unachievable. This paper presents a novel MPPI-based framework that enforces hard equality constraint satisfaction for closed-chain dual-arm manipulation. A Variational Autoencoder (VAE) learns a compact latent representation of the constraint manifold, enabling efficient near-feasible trajectory sampling. KKT-based projection is then applied in parallel across all samples on GPU, guaranteeing exact constraint satisfaction during trajectory optimization. A hierarchical control architecture decouples manifold-constrained trajectory optimization at 10 Hz from joint-level PD control at 1000 Hz, enabling real-time execution. Validated in MuJoCo simulation on a dual Franka Emika Panda platform, the proposed framework achieves stable constraint satisfaction and near real-time trajectory optimization across both goal-tracking and obstacle avoidance tasks.