Locomotion skills for simulated quadrupeds

Stelian Coros, Andrej Karpathy, Ben Jones, Lionel Revéret, Michiel van de PanneACM Trans. Graphics 2011 · DOI

TL;DR

This paper presents a comprehensive framework for simulating realistic quadrupedal locomotion in a physics-based environment. By introducing novel control abstractions like gait graphs and a flexible spine, and leveraging optimization techniques, the authors achieve a wide range of dynamic gaits, parameterized leaps, and essential behaviors like sitting and standing. This work is significant for its ability to generate robust, interactive, and visually convincing quadruped animations for applications in games, film, and robotics.

What Problem Does This Paper Address?

Animating realistic quadrupedal locomotion in a physics-based simulation is a challenging task. While kinematic approaches can produce visually appealing animations, they often lack the interactivity and responsiveness required for dynamic environments. Conversely, purely physics-based simulations can be difficult to control and author, especially for complex behaviors like varied gaits, transitions, and agile maneuvers. Existing methods often struggle with generating a diverse repertoire of motions, achieving robustness to external perturbations, and creating natural-looking movements that mimic real animals. The authors aim to bridge this gap by developing a unified control framework that combines the expressiveness of abstract representations with the robustness of physics-based simulation and the power of optimization.

Methods

The core of the paper lies in its novel control abstractions and their integration into a physics-based simulation. Key techniques include:

  • Gait Graphs: These graphs represent the timing and phasing of leg movements (swing and stance) for different gaits, providing a structured way to define periodic locomotion.
  • Dual Leg Frame Model: The quadruped is abstracted into front and rear "leg frames," each capable of independent foot placement, height, and pitch control. This simplifies the control problem by decoupling the control of different parts of the body.
  • Flexible Spine Model: A spine that can stretch and compress is introduced, which is hypothesized to contribute to more natural motion, particularly during fast gaits like the gallop, by allowing the body to remain more horizontal.
  • Internal Virtual Forces: These forces are applied via Jacobian transpose control to achieve desired motions without explicitly defining joint torques. This allows for abstract control of body posture and movement.
  • Optimization: A greedy stochastic local optimization algorithm is used to tune a large set of parameters (including trajectory parameters, stride durations, and virtual force gains) to achieve specific motion styles, robustness, and adherence to reference data.
  • Parameterized Leaps: A flexible leap controller is developed, parameterized by desired height, distance, and initial speed, allowing the quadruped to perform jumps onto platforms, over obstacles, and across gaps.
  • Behavioral Controllers: Separate controllers are designed for sitting, lying down, standing up, and getting up from a fall, often seeded by observing real-world animal behavior.

Datasets

The paper mentions the use of:

  • Experimental Video Data: Gait graphs for walk, trot, and pace were derived from tracking experimental video data of a dog.
  • Reference Motion Data: For evaluating the quality of simulated gaits, the authors used marker data from videos of real dogs, which they tracked and processed. This data was used to define the fd term in their objective function for optimization.

No new datasets were explicitly introduced in the paper.

Key Insights and Ideas

  • Abstraction is Key for Control: The paper demonstrates that abstracting complex quadrupedal motion into components like gait graphs and leg frames significantly simplifies the control design process, making it more manageable and scalable.
  • Flexible Spine for Naturalism: The flexible spine model appears to be crucial for achieving more natural and less oscillatory body movements, especially during dynamic gaits, by allowing for body extension and compression.
  • Optimization for Style and Robustness: Hand-tuning quadruped gaits is extremely difficult. The paper highlights the power of optimization in automatically discovering parameters that lead to robust gaits, specific motion styles, and adherence to desired performance metrics.
  • Unified Framework for Diverse Skills: The presented control abstractions are shown to be versatile enough to support not only periodic gaits but also complex, non-periodic skills like parameterized leaps and basic behavioral states.
  • Gait Transitions as Interpolation (Mostly): The paper shows that smooth gait transitions can often be achieved by interpolating gait parameters, though some transitions (like trot-to-canter) may require specialized controllers.

Key Results

  • Diverse Gait Repertoire: The system successfully simulates six distinct gaits: walk, trot, pace, canter, transverse gallop, and rotary gallop, with smooth transitions between them.
  • Robustness to Perturbations: The simulated gaits are shown to be robust to significant push disturbances (up to 400N for some gaits), demonstrating their stability in dynamic scenarios.
  • Parameterized Leaps: The quadruped can perform parameterized leaps with controllable height, distance, and speed, enabling it to traverse challenging terrain.
  • Behavioral Skills: The system can perform essential behaviors like sitting, lying down, standing up, and getting up from a fall.
  • Comparison to Real Data: Simulated gaits show reasonable correspondence with captured dog motion data, particularly in foot placements, though some differences in body and head motion are noted.
  • Efficiency: The simulation runs at 3x real-time on a current-generation PC, making it suitable for interactive applications.

Limitations and Future Work

  • Self-Collisions: The current implementation does not model self-collisions, which limits the flexibility of the spine and legs.
  • Scapula Modeling: The simplified modeling of the scapula at the shoulders might impact the fidelity of certain motions.
  • Turning Agility: While turning is demonstrated, the system is not yet as agile as real quadrupeds in this regard.
  • Foot Compliance: The simple box geometry for feet might be a limitation for highly agile motions requiring compliance.
  • Manual Design of Complex Behaviors: The get-up controller was manually designed, suggesting that optimization or more advanced authoring tools could be beneficial for such complex skills.
  • Future Work: The authors suggest extending the framework to a broader range of quadrupeds, incorporating more detailed biomechanical models (muscles, tendons), and developing more sophisticated methods for authoring and optimizing complex behaviors like landing on feet or pursuing prey. They also note the potential for improving turning agility and exploring more advanced foot models.

Summary updated

BibTeX
@article{coros2011locomotion,
  author = {Stelian Coros and Andrej Karpathy and Ben Jones and Lionel Revéret and Michiel van de Panne},
  title = {{Locomotion skills for simulated quadrupeds}},
  journal = {ACM Trans. Graphics},
  year = {2011},
  doi = {10.1145/2010324.1964954},
}
Figure