HCLM: A Hierarchical Framework for Cooperative Loco-Manipulation with Dual Quadrupeds

Technical Summary: HCLM – A Hierarchical Framework for Cooperative Loco-Manipulation with Dual Quadrupeds

Problem Statement
Real-world mobile manipulation requires integrating stable locomotion with precise object interaction. While single-agent systems have advanced, their kinematic reach limits spatially distributed interactions. Collaborative quadrupedal manipulators offer reconfigurable workspaces but face two primary challenges:

1. High-Level Coordination: Learning coordinate-free policies that allow spatially separated agents to execute diverse skills across varied configurations, rather than relying on pre-defined actions.
2. Low-Level Execution: Achieving stable whole-body control that guarantees precise task-space tracking, rejects floating-base disturbances, and strictly regulates internal forces within closed kinematic chains to prevent destructive stresses.

Existing approaches often fail to meet these requirements, being limited to cooperative locomotion without manipulation, non-prehensile interactions (e.g., pushing), or specific pre-defined transportation tasks lacking scalability and fine-grained 6D pose control.

Methodology: The HCLM Framework
The authors propose HCLM, a hierarchical framework that systematically decouples high-level collaborative reasoning from low-level robust motion execution for dual quadrupedal systems.

1. High-Level: Canonicalized Joint Diffusion Policy

   • Architecture: A centralized imitation learning module using a Joint Diffusion Policy.
   • Representation: The policy operates within an SE(3)-invariant task space. Instead of using absolute global coordinates, it anchors the task reference frame to the base of the primary robot at the initial time step. All inputs (end-effector poses) and outputs (target poses) are transformed into this relative frame.
   • Benefit: This formulation mathematically ensures spatial invariance, allowing the policy to learn coordinate-agnostic coordination patterns independent of global pose estimation drift. It processes concatenated state histories of both agents to infer synchronized end-effector trajectories.

2. Low-Level: Task-Centric Hybrid Whole-Body Controller (WBC)
   The WBC translates frame-agnostic references into physical motion via two synergistic layers:

   • Proactive Layer (Kinematic MPC): A Model Predictive Control (MPC) module operates over a receding horizon to compute optimal, collision-free velocity distributions for the floating bases and manipulators. It acts as a reference generator, explicitly delegating dynamic stabilization to the reactive layer to avoid computational prohibitions of full nonlinear dynamics on edge devices.
   • Reactive Layer (Hybrid Execution):
     • Base Stabilization: A Reinforcement Learning (RL) policy, trained with domain randomization (randomized payloads and external wrenches), robustly tracks the base velocities commanded by the MPC.
     • Manipulator Control & Force Regulation: A high-frequency Quadratic Programming (QP) solver computes instantaneous joint velocities. Crucially, this layer integrates a cooperative admittance scheme. It uses a grasp matrix to orthogonally decompose contact wrenches into external (motion-inducing) and internal (object-squeezing) forces. Internal force errors are mapped to corrective yield velocities, relaxing rigid positional tracking to safely dissipate kinematic conflicts and regulate internal stresses during closed-chain interactions.

Key Contributions

1. HCLM Framework: A hierarchical system enabling dual quadrupedal manipulators to seamlessly integrate stable locomotion with precise object interaction for complex, spatially distributed tasks.
2. SE(3)-Invariant Policy & Hybrid WBC: The introduction of a frame-agnostic task-space representation for learning cooperative policies, executed by a hybrid controller that synergizes proactive motion distribution with reactive stabilization and active force regulation.
3. Validation: Comprehensive evaluation across three progressively challenging tasks (Dynamic Handover, Cooperative Carrying, Cooperative Packing) in simulation and a real-world deployment of the Dynamic Handover task.

Experimental Results

• Configuration Agnosticism: In simulation, the SE(3)-invariant formulation (BW) achieved a 62.5% success rate (50/80 trials) under arbitrary global orientations. In contrast, a baseline policy using absolute world coordinates (W) suffered catastrophic failure (0/80) under the same conditions, demonstrating the critical necessity of the invariant representation for generalization.
• Robustness to Perturbations: The system maintained reliable performance (success rate drop within 10–25%) under external base perturbations up to 100 N. The end-effector deviation remained tightly bounded (<2 cm) despite significant base displacement (~8 cm), validating the WBC's ability to isolate base instabilities from task execution.
• Closed-Chain Compliance: Without the cooperative admittance scheme, spatial drifts induced destructive internal forces (>50 N), causing task failure. With the scheme, the system successfully dissipated antagonistic forces, maintaining stability.
• Real-World Deployment: The framework was successfully deployed on physical hardware (two Unitree Go2 bases with ARX5 arms) for a dynamic handover task, demonstrating the transferability of the SE(3)-invariant policy from simulation to reality.

Significance
The paper claims that HCLM offers a highly robust pathway for multi-robot embodied coordination. By decoupling high-level spatial reasoning from low-level dynamic execution and enforcing coordinate invariance, the framework addresses the scalability and robustness gaps in current multi-robot manipulation systems. It enables dual quadrupedal systems to perform complex, general-purpose cooperative tasks with exceptional resilience against physical perturbations and kinematic inconsistencies, moving beyond pre-defined scenarios to adaptive, distributed manipulation.