Co-designing versatile quadruped robots for dynamic and energy-efficient motions

Gabriele Fadini, Shivesh Kumar, Rohit Kumar, Thomas Flayols, Andrea Del Prete, Justin Carpentier, Philippe SouèresRobotica 2024 · DOI

In traditional robotics, hardware design and controller development are often siloed: engineers build a robot, and then programmers attempt to optimize its movement. However, a robot's physical "blueprint"—such as the relative lengths of its limbs—fundamentally dictates its performance limits. This paper, published in Robotica (2024), presents a robust "co-design" framework that simultaneously optimizes a quadruped’s physical dimensions and its motion control to achieve peak energy efficiency.

Bridging the Gap in Robot Design

While the concept of co-design is not entirely new, previous attempts often relied on oversimplified models that ignored real-world hardware constraints like motor friction, battery voltage limits, and actuator bandwidth. This research bridges that gap by incorporating high-fidelity dynamics and hardware-specific limitations into the optimization process.

Comparison of the proposed framework's features against existing literature.

As shown in the comparison above, this work distinguishes itself by integrating full system dynamics with hard constraints, cyclicity, and specifically, actuator bandwidth—a factor often overlooked that prevents the optimizer from suggesting physically impossible torque changes.

The Bi-Level Optimization Framework

The researchers developed a bi-level optimization scheme to manage the complexity of simultaneous hardware and motion tuning.

The bi-level optimization framework used for robot co-design.
  1. Outer Loop (Hardware): A genetic algorithm (CMA-ES) explores the design space, selecting link scaling factors (thigh, shank, and base lengths) and discrete motor types.
  2. Inner Loop (Motion): For each design candidate, a trajectory optimization problem is solved using IpOPT. This finds the most energy-efficient way for that specific hardware configuration to perform tasks like bounding or jumping.

Key Innovations

The framework introduces several critical modeling improvements:

  • Actuator Modeling: By including a first-order filter for actuator bandwidth, the model ensures that torque changes remain within physical limits.
  • Cyclicity Constraints: The system enforces "non-Markovian" constraints, ensuring the robot ends a motion in a state ready to begin the next cycle, which is vital for continuous locomotion.

Results: Efficiency and Versatility

The co-designed "optimal" robot demonstrated significant performance gains, showing an improvement of at least 52% in energy efficiency for both bounding and backflip tasks compared to initial nominal designs.

The framework's reliability was validated using technical specifications from Antigravity AK series motors and experimental data from a prototype at the DFKI Robotics Innovation Center. Measured power consumption (39.69J) almost perfectly matched the model's prediction (39.64J).

The study also revealed that the "ideal" robot shape is task-dependent. For forward jumping, the optimizer favored longer thighs, whereas longer shanks were superior for backflips to help generate angular momentum. To address this, the authors used a Pareto set to identify a "sweet spot" design (Thigh scaling 0.66, Shank scaling 1.03) that balances performance across multiple tasks.

Conclusion and Future Work

While the current optimization process is computationally intensive—taking roughly 10 hours on a high-performance cluster—it provides a clear blueprint for the next generation of versatile, energy-efficient quadruped robots. Future work will focus on automating gait discovery and physically fabricating the optimized designs.

Summary updated

BibTeX
@article{fadini2024co,
  author = {Gabriele Fadini and Shivesh Kumar and Rohit Kumar and Thomas Flayols and Andrea Del Prete and Justin Carpentier and Philippe Souères},
  title = {{Co-designing versatile quadruped robots for dynamic and energy-efficient motions}},
  journal = {Robotica},
  year = {2024},
  doi = {10.1017/s0263574724000730},
}
Figure