Predictive Control with Indirect Adaptive Laws for Payload Transportation by Quadrupedal Robots

This paper presents a novel hierarchical control framework that integrates an indirect adaptive law with model predictive control to enable quadrupedal robots to robustly transport heavy static and dynamic payloads across diverse terrains by estimating unknown parameters and ensuring stability through a convex stability criterion.

The Gist

Imagine a four-legged robot dog trying to carry a heavy, mysterious box across a bumpy, rocky path. The problem is, the robot doesn't know exactly how heavy the box is, or where its center of gravity is. If the robot guesses wrong, it might trip, drop the box, or fall over.

This paper presents a new "brain" for these robots that solves this problem using a clever two-step strategy: Learning while walking and Planning ahead with safety nets.

Here is how it works, broken down into simple concepts:

1. The Two-Level Brain (The Architect and the Muscle)

Think of the robot's control system as having two distinct roles:

• The High-Level Architect (The "Big Picture" Planner): This part of the brain doesn't worry about every tiny muscle twitch. Instead, it looks at the robot as a single, simplified block (like a heavy suitcase). Its job is to figure out the path forward and, crucially, to guess the weight of the box in real-time.
• The Low-Level Muscle (The "Doer"): This is the detailed controller that actually moves the robot's 18 joints. It takes the simple path from the Architect and figures out exactly how much force to apply to each leg to stay balanced.

2. The "Guess and Check" Game (Indirect Adaptive Law)

The biggest challenge is that the robot doesn't know the weight of the payload. The old way was to guess once and hope for the best. This new method is like a smart detective:

• As the robot walks, the "Architect" constantly watches how the robot reacts to the ground.
• If the robot sinks a little lower than expected, the Architect thinks, "Ah, I guessed the weight was 10kg, but it's acting like 15kg. I need to update my guess."
• It uses a mathematical tool called Gradient Descent (think of it as sliding down a hill to find the lowest point) to quickly adjust its estimate of the weight and the box's balance.
• The Magic Trick: The paper proves mathematically that this guessing game won't go crazy. It guarantees that the robot's guess will get closer and closer to the truth, never spiraling out of control.

3. The Safety Net (Convex Stability)

Usually, when you try to make a robot learn while it moves, it's like trying to juggle while riding a unicycle on a tightrope. If the robot's guess is slightly off, it could fall.

The authors added a special safety rule (a "convex stability criterion") into the planning algorithm.

• Analogy: Imagine the robot is driving a car. The "safety rule" is like a guardrail that says, "You can steer anywhere you want, but you must stay within these specific lines where we know the car won't flip over, even if our weight estimate is slightly wrong."
• This ensures that even while the robot is learning the weight of the box, it never plans a move that would cause it to crash.

4. The Results: Carrying Heavy Loads on Rough Ground

The team tested this on a Unitree A1 robot (a small, agile quadruped).

• The Test: They strapped heavy, unknown weights to the robot's back and sent it over wooden blocks, grass, and gravel.
• The Record: The robot successfully carried loads up to 109% of its own body weight on flat ground and 91% on rough, rocky ground.
• The Comparison: When they compared this new "smart learning" brain to standard robot brains (which don't learn on the fly) or other adaptive methods, the new system was far superior. It didn't just survive; it trotted confidently.

Why This Matters

In the real world, robots need to help humans carry things. But humans don't always carry the same weight, and the ground isn't always flat.

• Old Robots: "I think this box is light. Crash."
• This New Robot: "Hmm, I'm sinking a bit. Let me recalculate the weight... okay, it's heavier. Adjusting my steps and balance... Success!"

In short, this paper gives robots the ability to adapt on the fly, turning them from rigid machines that break under surprise loads into resilient helpers that can carry heavy, unknown objects through the messiest environments imaginable.

Technical Summary

Here is a detailed technical summary of the paper "Predictive Control with Indirect Adaptive Laws for Payload Transportation by Quadrupedal Robots."

1. Problem Statement

Quadrupedal robots operating in human-centric environments (factories, homes) must navigate uneven terrains while carrying unknown and varying payloads. This presents a significant control challenge due to:

• Model Uncertainty: The mass and moment of inertia of the robot change dynamically when carrying payloads, which are often unmodeled and unknown.
• System Complexity: Full-order locomotion models are high-dimensional, nonlinear, and hybrid, making real-time control difficult.
• Limitations of Existing Methods: Standard Model Predictive Control (MPC) relies on accurate prior knowledge of system parameters. Existing adaptive approaches often lack formal stability guarantees for estimation errors or are computationally too heavy for real-time implementation on complex terrains.

The core problem addressed is developing a real-time, hierarchical control framework that can robustly transport unknown static and dynamic payloads on rough terrains while formally guaranteeing the stability of the parameter estimation process.

2. Methodology

The authors propose a Hierarchical Adaptive Model Predictive Control (AMPC) framework integrated with a Low-Level Whole-Body Controller (WBC).

A. High-Level: Adaptive MPC (AMPC)

• Template Model: The system utilizes a Single Rigid Body (SRB) model (a reduced-order "template") to represent the robot's centroidal dynamics. This reduces computational complexity compared to full-order models.
• Indirect Adaptive Law:
  • Instead of directly estimating the full system state, the controller estimates the uncertain parameters (mass and inertia properties) of the SRB model.
  • A gradient-descent-based updating law is employed to estimate these parameters ($\hat{\theta}$) in real-time. The update rule is: $\hat{\theta}(t + 1) = \hat{\theta}(t) + \lambda \Gamma^\top(t) \tilde{x}(t + 1)$, where $\tilde{x}$ is the state estimation error and $\Gamma$ is a regressor matrix.
• Stability Guarantees:
  • The paper derives a sufficient condition for the asymptotic stability of the parameter estimation error.
  • Crucially, the authors transform this non-convex stability condition (involving eigenvalues) into a convex inequality constraint (Lemma 1). This allows the stability criterion to be embedded directly into the Quadratic Programming (QP) formulation of the MPC.
• Optimization: The AMPC solves a convex QP at 160 Hz to generate optimal reduced-order trajectories (COM position, velocity, orientation) and Ground Reaction Forces (GRFs) that satisfy the convex stability constraints.

B. Low-Level: Nonlinear Whole-Body Controller (WBC)

• Tracking: The optimal trajectories and GRFs from the high-level AMPC are passed to a low-level controller running at 1 kHz.
• Mechanism: This controller uses a QP-based approach with virtual constraints (feedback linearization) to track the desired trajectories. It computes joint torques and ensures no-slip conditions for the stance feet.
• Architecture: This separation allows the high-level planner to handle uncertainty and planning, while the low-level controller handles the complex dynamics and contact constraints of the full 18-DOF robot.

3. Key Contributions

1. Novel Indirect Adaptive Law: Development of a gradient-descent-based parameter estimation algorithm specifically designed for reduced-order locomotion models, capable of handling unknown payloads.
2. Formal Stability Integration: Theoretical proof of the asymptotic stability of the estimation error, with the stability condition successfully embedded as a convex constraint within the MPC framework. This is a significant advancement over previous methods that either lacked formal guarantees or were non-convex.
3. Hierarchical Framework: A seamless integration of high-level adaptive planning (AMPC) and low-level tracking (WBC) that bridges the gap between reduced-order templates and full-order dynamics.
4. Robustness to Extreme Uncertainty: The method is designed to handle payloads significantly larger than the robot's own mass, including dynamic (time-varying) additions and external disturbances.

4. Experimental Results

The framework was validated through extensive numerical simulations and hardware experiments on the Unitree A1 quadruped robot (12.45 kg).

• Payload Capacity:
  • Simulations: Successfully transported payloads up to 132% of the robot's mass.
  • Flat Terrain Experiments: Successfully carried 109% of its mass (13.6 kg).
  • Rough Terrain Experiments: Successfully carried 91% of its mass (11.34 kg) on wooden blocks and uneven ground. This is a record for model-based control on rough terrain for this robot class (previous works reported ~50% on rough terrain).
• Dynamic Scenarios: The robot maintained stability when dynamic payloads were added mid-walk (up to 73% uncertainty) and when subjected to external push disturbances.
• Comparative Performance:
  • Success Rate: On 1,500 randomly generated rough terrains, the proposed AMPC achieved an 88.22% success rate, compared to 18.89% for normal MPC and 67.31% for an L1 Adaptive MPC.
  • Trajectory Tracking: The AMPC maintained desired velocities and heights significantly better than standard MPC, especially under high payload loads (e.g., 10 kg).
• Real-Time Performance: The AMPC solver runs in ~0.95 ms, and the WBC in ~0.21 ms, confirming feasibility for real-time deployment.

5. Significance

This work represents a major step forward in the deployment of legged robots for practical logistics and rescue missions where payload weights are unknown and terrains are unstructured.

• Theoretical Rigor: By embedding stability guarantees directly into the convex MPC constraints, the paper moves beyond heuristic adaptive control to a mathematically rigorous framework.
• Practical Impact: The ability to carry payloads nearly equal to or exceeding the robot's own weight on rough terrain demonstrates a level of robustness previously unattainable with standard model-based controllers.
• Scalability: The hierarchical approach (Template + Full-Order) provides a scalable architecture that can be extended to other legged robots and more complex hybrid dynamics (e.g., incorporating leg switching logic in future work).

In summary, the paper presents a robust, theoretically grounded control solution that enables quadrupedal robots to adapt to extreme physical uncertainties in real-time, significantly expanding their operational envelope in real-world scenarios.