TL;DR
This paper tackles the challenging problem of transitioning a legged robot from flat ground to vertical wall climbing. They propose a novel motion planning approach using complementarity constraints to automatically discover optimal contact sequences, enabling robots to autonomously navigate this complex maneuver.
What Problem Does This Paper Address?
Autonomous vertical wall climbing is a crucial capability for robots operating in unstructured environments like disaster sites or industrial maintenance. A significant hurdle is the "transition phase" – the maneuver of moving from stable ground contact to clinging to a vertical surface. This phase is complex because it requires simultaneous planning of multiple contacts and the forces (wrenches) exerted at those contacts. Traditional methods often rely on pre-defined contact sequences, which become computationally intractable as the number of possible sequences explodes with the robot's limbs. Furthermore, real-world factors like imprecise friction coefficients and motor torque limitations make robust planning difficult. This paper aims to develop a method that can autonomously plan this transition, considering these real-world constraints and discovering flexible contact sequences without human pre-specification.
Methods
The core of the proposed method is a nonlinear programming (NLP) formulation that leverages complementarity constraints. These constraints are key to handling the discrete nature of contact (either a foot is in contact or it's not) within a continuous optimization framework. Specifically, the paper uses complementarity between:
- Contact distance and normal contact force: This ensures that contact forces are only generated when a foot is at zero distance to the surface, and vice-versa. This allows the planner to explore various contact sequences without explicitly enumerating them.
- Limb switchability (optional): This constraint helps mitigate local minima by encouraging the robot to switch supporting limbs over time, preventing prolonged reliance on a single limb which could lead to motor overheating.
The planner also incorporates:
- Static equilibrium constraints: Ensuring the robot remains stable under gravity and contact forces.
- Kinematic reachability constraints: Limiting foot placements to within the robot's physical capabilities.
- Safety factors: Two crucial parameters are introduced:
- $S_\mu$: A factor to account for uncertainty in the friction coefficient, making the plan more conservative on slippery surfaces.
- $S_\tau$: A factor to account for motor torque limitations, preventing over-torquing.
These safety factors are tunable and directly influence the generated contact sequences, allowing the planner to adapt to different environmental conditions and robot limitations. The optimization problem is solved using SNOPT, a nonlinear programming solver. For hardware implementation, an indirect force control method is used to achieve the desired contact forces on the position-controlled robot.
Datasets
Not discussed in the paper. The paper focuses on simulation and hardware demonstrations of a specific robot (SILVIA).
Key Insights and Ideas
- Complementarity as a powerful tool for contact planning: By formulating contact as a complementarity problem, the planner can implicitly explore a vast number of contact sequences, overcoming the combinatorial explosion of traditional discrete search methods.
- Tunable safety factors for real-world robustness: The introduction of $S_\mu$ and $S_\tau$ provides a principled way to incorporate real-world uncertainties (friction, motor limits) into the planning process, making the generated motions more likely to succeed on hardware.
- Flexibility in contact sequences: The method demonstrated that it can generate diverse contact sequences (e.g., 3-3, 4-2, 2-2-2) depending on the environmental setup and safety factor tuning, showcasing its adaptability.
- Importance of contact torque (even if not directly controlled): While the robot's end-effectors don't directly control torque, the paper highlights that considering contact torque can improve stability, suggesting future work in this area.
Key Results
The paper successfully demonstrates the proposed transition motion planner on the SILVIA hexapod robot. Key results include:
- Successful transitions: The robot was able to transition from flat ground to vertical walls in various scenarios, including parallel walls, parallel walls with steps, and even a circular tube.
- Generation of diverse contact sequences: The planner autonomously generated different contact sequences (e.g., 3-3, 4-2, 2-2-2, 3-3-3, 3-1-2-2) based on environmental conditions and safety factor tuning. For instance, a 4-2 sequence was generated for a more slippery wall, and a 3-1-2-2 sequence was found to be robust for navigating steps.
- Hardware validation: The generated motion plans were implemented on the SILVIA robot, validating the feasibility of the approach in real-world experiments.
- Efficient computation: The NLP solver (SNOPT) typically generated motions within 10 seconds, with specific parallel wall tasks averaging between 1.054s and 3.083s on an Intel Core i7 CPU.
Limitations and Future Work
- Pre-determined planning horizon (M): The number of planning rounds (M) is currently set based on human experience. Future work could involve determining M automatically.
- Contact torque consideration: While the paper acknowledges the importance of contact torque for stability, the current formulation doesn't explicitly control it due to hardware limitations. Future work could explore how to better incorporate this.
- Friction coefficient estimation: The planner relies on prior knowledge of the friction coefficient. Future work could integrate real-time estimation of friction using vision or other sensors.
- Online execution: While current running times are reasonable, the authors aim to accelerate the planner with C++ implementation for true online execution.
- Exploration of more complex environments: Extending the planner to handle more dynamic or unpredictable environments would be a valuable next step.
Summary updated