TL;DR
This paper introduces a novel framework for dynamic jumping in bipedal robots, addressing the challenge of unstable landings. Their key innovation is a heuristic landing planner that uses real-time momentum feedback during the flight phase to dynamically adjust landing locations, significantly improving stability and robustness.
What Problem Does This Paper Address?
Dynamic jumping is a crucial capability for legged robots to achieve human-like mobility, but it presents significant control challenges. The primary difficulties lie in the uncontrollable "floating base" during the flight phase and the substantial impact forces upon landing. Bipedal robots, with their inherently smaller and less stable footprints compared to multi-legged robots, are particularly susceptible to these issues. Existing methods often struggle to generate complex jumping motions (like twists or somersaults) or to ensure safe and stable landings, especially when faced with uncertainties from tracking errors or external disturbances. This paper aims to bridge this gap by developing a comprehensive framework that enables robust and versatile dynamic jumping for bipedal robots.
Methods
The proposed framework is a hierarchical system combining several key components:
- Kino-dynamics Motion Planner with Centroidal Momentum: This high-level planner generates the desired jumping trajectory offline. It utilizes centroidal dynamics, which accurately capture the robot's full-body mass and inertia distribution, allowing for more natural and compensatory body rotations to manage angular momentum generated by leg movements. This is a departure from simpler point-mass models.
- Heuristic Landing Planner: This is the paper's core novelty. During the flight phase, it uses real-time feedback of the robot's linear and angular momentum to predict and update the desired landing locations for the feet. This allows the robot to react to deviations from the planned trajectory or external disturbances, minimizing landing impact. It's inspired by human subconscious reactions to maintain balance.
- Low-Level Whole-Body Controller: This controller, implemented as a quadratic programming (QP) problem, tracks the modified trajectory (including the heuristically adjusted landing points) at a high frequency (500 Hz). It prioritizes tasks like maintaining COM position and velocity, controlling torso orientation, and damping excessive angular momentum.
Datasets
Not discussed in the paper. The paper focuses on simulation and hardware experiments using their developed robot, BRUCE.
Key Insights and Ideas
- Leveraging the Flight Phase for Landing: The most significant insight is the realization that the flight phase, often considered uncontrollable, can be actively used to improve landing stability. By using momentum feedback, the robot can "correct" its landing trajectory in mid-air.
- Centroidal Dynamics for Natural Motion: Using centroidal dynamics allows the planner to generate more human-like jumping motions where the robot actively rotates its body to compensate for leg movements, rather than treating the body as a simple point mass.
- Heuristic Landing as a Robustness Enabler: The heuristic landing planner provides a fast and effective way to handle uncertainties, which is crucial for bipedal robots with limited stability margins. This is a practical approach that can be implemented in real-time.
- Hierarchical Control for Versatility: The combination of a high-level planner for complex motion generation and a low-level controller for precise tracking and adaptation is key to achieving a wide range of dynamic behaviors.
Key Results
The framework was demonstrated on a miniature bipedal robot named BRUCE.
- Versatile Jumping: The robot successfully performed various jumping tasks, including in-situ jumps, directional jumps, twisting jumps, and step jumps.
- Somersaults: The framework enabled the robot to perform a somersault in simulation and, with relaxed torque limits, even in-situ.
- Robustness: The heuristic landing planner showed effectiveness in push recovery simulations, allowing the robot to land safely even after being subjected to external forces during the flight phase.
- Performance Metrics: Maximum jumping height was around 15 cm (22.7% of total height), maximum jumping distance was about 10 cm, and twisting angles reached up to 30 degrees.
Limitations and Future Work
- Actuator Limitations: The current hardware has limitations in torque, which prevented certain dynamic maneuvers like in-situ somersaults without relaxation.
- Body Inertia: The body's inertia is not large enough compared to the leg inertia, leading to body rotation during leg movements that the landing planner has to compensate for.
- Planner Speed: The offline motion planner can have a long solving time (up to 5 minutes), which is being addressed by exploring offline motion libraries for faster online execution.
- Future Work: The authors suggest upgrading the robot with a more optimized mass distribution to further enhance performance and exploring offline motion libraries for faster dynamic behavior execution.
Summary updated