TL;DR
This paper presents a novel framework for quadrupedal robots to navigate challenging, unknown discrete terrains using visual feedback. By combining a pre-computed library of dynamic gaits with a geometrically-aware model predictive controller and real-time visual perception, the robot can robustly step on discrete footholds.
What Problem Does This Paper Address?
Quadrupedal robots are adept at traversing rough terrain, but navigating environments with discrete, uneven footholds (like stepping stones) presents a significant challenge. This is because the robot's motion can become highly dynamic and unpredictable, requiring precise foot placement while maintaining stability. Traditional methods often struggle with the high dimensionality of the problem, real-time constraints, and the need to reason about both feasible footholds and desired robot velocities (to speed up or slow down). Furthermore, the robot's orientation dynamics on challenging terrain can lead to instability, and existing control methods often rely on simplifications that limit their applicability. This paper aims to bridge this gap by enabling robust, vision-guided locomotion on such challenging terrains.
Methods
The core of the proposed framework lies in a combination of offline gait generation and online control:
- Offline Motion Library Generation: The authors build a library of "two-step" periodic trotting gaits. This involves pre-computing a set of gaits using trajectory optimization, where each gait is parameterized by step lengths. The "two-step" periodicity allows for more flexibility in choosing step lengths independently compared to "one-step" periodic gaits, which is crucial for adapting to varying terrain. This library serves as a reference for desired robot motion.
- Geometric Model Predictive Control (GVMPC): For online control, a novel Geometric MPC is introduced. This controller operates on a reduced-order rigid-body model of the quadruped. Crucially, it employs a variation-based linearization technique that preserves the geometric structure of the robot's orientation dynamics on the SO(3) manifold. This approach avoids singularities and ensures energy conservation, making it more robust than methods relying on small-angle approximations, especially on uneven terrain. The GVMPC outputs contact forces for the stance legs.
- Vision-Aided Localization and Mapping: To perceive the environment, the system uses a forward-facing depth camera. A contact-aided invariant Extended Kalman Filter (EKF) provides robot localization, and a probabilistic mapping framework generates a local height map of the terrain. This map is used to identify feasible foothold locations, particularly for the hind legs, which are otherwise difficult to perceive.
Datasets
Not discussed in the paper. The experiments were conducted on custom-built discrete terrains (aligned, staggered, and random) using a Unitree A1 quadruped robot.
Key Insights and Ideas
- "Two-step" periodicity for gait flexibility: The insight that "two-step" periodic gaits offer sufficient flexibility to independently choose step lengths for quadrupedal locomotion is key. This allows for more adaptable gaits than simpler "one-step" periodic models.
- Geometric MPC for robust orientation control: Using a geometrically-aware MPC that linearizes dynamics on the SO(3) manifold is a significant advancement. This addresses the limitations of traditional methods that struggle with large orientation changes and potential singularities, leading to more stable locomotion on challenging terrain.
- Vision for hind-limb footstep planning: The integration of vision for mapping and localization is crucial for enabling hind-limb footstep planning. This overcomes the inherent limitation of forward-facing sensors in perceiving terrain behind the robot.
- Combining offline and online strategies: The synergy between a pre-computed, diverse gait library and a reactive, geometrically-sound online controller is a powerful combination for dynamic locomotion.
Key Results
- The proposed GVMPC framework successfully enabled a Unitree A1 quadruped robot to navigate multiple unknown, challenging discrete terrains in real-world experiments.
- The approach outperformed two baseline controllers: a heuristic method and a Jacobian-linearized model with a gait library.
- Success rates on aligned and staggered terrains were significantly higher with the proposed method compared to baselines.
- The system demonstrated robustness on random terrains, which require more precise foot placement and CoM control.
Limitations and Future Work
- Terrain segmentation: The method requires the elevation map to be segmented into stepable and un-stepable regions, which is currently done via thresholding. Inaccuracies can arise from depth image occlusions and estimation drift.
- Localization drift: Any drift in the robot's position estimate can lead to inaccurate foot placement. Tuning the invariant EKF for the compliant feet and contact sensors proved challenging.
- Future Work: The authors suggest extending the framework to handle varying step widths and turning maneuvers on discrete terrain. They also note that learning-based approaches could potentially replace the current localization and mapping framework for more adaptive foothold selection.
Summary updated