Capability-Aware Planning

A path planner that ignores what its robot can physically do produces plans that fail at execution time. A quadrupled step height, a topology change from six legs to four, a different low-level controller — each changes which terrains are traversable. I formalize this as the robot’s traverse capability: the probability that a robot, with its current morphology and controller, successfully crosses a given piece of terrain.

Approach

My work turns capability into a first-class, learned quantity with three components:

1. Capability maximizing — reinforcement learning selects the control strategy and body topology that maximize traverse capability for the terrain at hand.
2. Capability abstraction — a supervised network predicts the robot’s capability from terrain and foothold state, trained on large numbers of simulated locomotion trials.
3. Capability-based planning — classic path planners are augmented with the learned capability model, so long-range guidance paths conform to what the robot can execute.

Evidence

The CapPlanner system (ROBIO 2022, Best Paper in Biomimetics finalist) and its journal extension demonstrate the full pipeline on the Qingzhui hexapod: long-range locomotion experiments in simulation and the real world show large improvements in global locomotion success rate compared to capability-blind planning.

Where this is going

Capability models are a natural interface between perception, control, and planning — a compact summary of “what my body can do here.” I want to extend them to dynamic skills, degraded hardware (missing legs, weak motors), and as priors for exploration in new environments.