TL;DR
Researchers have developed PACC (Passive-Arm for Collaborative Carrying), a minimalist, unactuated mechanical arm that enables quadruped robots to carry heavy loads collaboratively over rough terrain. By replacing heavy active manipulators with a lightweight spring-damper system and a decentralized Model Predictive Controller (MPC), the system achieves stable human-robot and robot-robot coordination without the complexity or weight of traditional robotic arms.
What Problem Does This Paper Address?
Collaborative carrying (CC) is essential for logistics in unstructured environments, such as transporting construction materials or rescue victims. While legged robots are ideal for this, current approaches often rely on active manipulators. These actuated arms are problematic: they are expensive, fragile, and heavy—sometimes consuming over 40% of the robot's total payload capacity just to carry the arm itself.
Furthermore, rigid mechanical couplings between robots create "holonomic constraints" that are difficult to manage on uneven ground. If one robot needs to sway to maintain balance on a rock, a rigid link forces that motion onto the other robot, potentially destabilizing both. The authors seek a "minimalist" solution that maximizes net payload and handles the unpredictable nature of rough terrain and human leaders.
Methods
The authors propose a two-part solution involving a novel hardware design and a specialized control scheme:
1. The PACC Hardware
The PACC is a 3-degree-of-freedom (DoF) passive arm (yaw-pitch-pitch) weighing only 1.3 kg. Instead of motors, it uses:
- Springs and Damping: These provide intrinsic impedance, allowing the arm to act like a compliant pendulum.
- Encoders: These measure joint displacements, which serve as the primary sensor for understanding the leader's movement and the forces exerted by the payload.
2. Locomotion Guidance and Control
The system uses a decentralized approach, meaning each robot makes its own decisions based on local "proprioceptive" data:
- Haptic Guidance: The follower robot treats the arm as a joystick. The angular displacement of the joints ($\theta$ for pitch, $\psi$ for yaw) determines the desired forward and heading velocities.
- Force-Aware MPC: The controller uses a Single Rigid Body Dynamics (SRBD) model. It estimates external forces at the end-effector using the arm's spring model and incorporates these into the MPC. This allows the robot to proactively adjust its trunk sway and footholds to keep its Zero-Moment Point (ZMP) stable despite the shifting weight of the payload.
Datasets
The researchers did not use standard datasets; instead, they validated the system through real-world experimental scenarios using two Unitree Aliengo quadruped robots (21kg each). The testing ground was a "two-stage circuit" consisting of:
- Stair-like obstacles: Pallets ranging from 13cm to 16cm in height.
- Rough terrain: A bed of irregular rocks.
Key Insights and Ideas
- Compliance as a Feature: By using a passive arm, the robots are mechanically decoupled. This allows each robot to adjust its body attitude and footholds independently on rough terrain without fighting the other robot through a rigid payload.
- The Pendulum Advantage: The third link of the arm is designed to behave like a pendulum. This makes the oscillatory dynamics of the payload predictable and independent of the payload's mass, simplifying force estimation for the MPC.
- Minimalism Wins: By removing actuators, the researchers increased the robot's effective payload capacity and made the system more robust to falls, as there are no expensive gearboxes to shatter upon impact.
Key Results
The system was validated across three distinct scenarios:
- Robot-Robot (Rigid Payload): Two robots carried a 7kg rigid bar across stairs and rocks. The MPC successfully compensated for pendulum-like oscillations (approx. 1s period) to maintain stability.
- Robot-Robot (Deformable Payload): The robots carried a 2kg load via a 60cm rope. The controller maintained enough tension in the rope to keep the payload off the ground while traversing obstacles.
- Human-Robot CC: A human acted as the leader, carrying a 7kg rigid payload with a robot follower. The robot successfully interpreted the human's intent through the passive arm's displacement, even on challenging terrain.
Limitations and Future Work
While successful, the authors noted several areas for improvement:
- Blind Locomotion: The current system does not use vision for foothold selection, relying entirely on the controller's reactivity. Integrating vision-based terrain mapping would improve traversability.
- Design Optimization: The current PACC is a preliminary prototype; further optimization of the spring constants and link lengths could enhance performance for specific payload classes.
- Complex Maneuvers: Because the arm is passive, it cannot perform grasping or active lifting. It is strictly a transportation tool.
Summary updated