Designing Tools with Control Confidence

TL;DR

The paper tackles robust autonomous hand-tool design for robots by introducing a control-confidence term inspired by metacognition and the Free Energy Principle to address perturbations in tool use. It formalizes a parametric tool design problem with a free-energy objective $F=\frac{1}{2}(X-X^g)^T P^X (X-X^g)-\frac{1}{2}\log|\Pi^U|$, where $E(t)=\frac{1}{2}(X(t)-X^g)^T P^X (X(t)-X^g)$ and $\Pi^U=\frac{\partial^2 E}{\partial U^2}$, and optimizes using CMA-ES with $P^X=p^X I_3$. Simulation results in PyBullet with a 7-DOF Franka arm show that tools designed with higher control confidence are more robust to environmental perturbations (e.g., box mass changes), while the free-energy objective balances robustness and goal accuracy; CMA-ES outperforms other optimizers in design speed. The work thus enables metacognitive, uncertainty-aware tool design with practical implications for robust robotic manipulation, and provides code for replication.

Abstract

Prehistoric humans invented stone tools for specialized tasks by not just maximizing the tool's immediate goal-completion accuracy, but also increasing their confidence in the tool for later use under similar settings. This factor contributed to the increased robustness of the tool, i.e., the least performance deviations under environmental uncertainties. However, the current autonomous tool design frameworks solely rely on performance optimization, without considering the agent's confidence in tool use for repeated use. Here, we take a step towards filling this gap by i) defining an optimization framework for task-conditioned autonomous hand tool design for robots, where ii) we introduce a neuro-inspired control confidence term into the optimization routine that helps the agent to design tools with higher robustness. Through rigorous simulations using a robotic arm, we show that tools designed with control confidence as the objective function are more robust to environmental uncertainties during tool use than a pure accuracy-driven objective. We further show that adding control confidence to the objective function for tool design provides a balance between the robustness and goal accuracy of the designed tools under control perturbations. Finally, we show that our CMAES-based evolutionary optimization strategy for autonomous tool design outperforms other state-of-the-art optimizers by designing the optimal tool within the fewest iterations. Code: https://github.com/ajitham123/Tool_design_control_confidence.

Figures (9)

• Figure 1: Designing tools for high controllability: a) Depicts a prehistoric human performing a manipulation task with a stick tool, b) Our PyBullet simulation environment mimicking (a) with a manipulator arm and a tool to manipulate a box on the ground, c) The final results of our tool design algorithm for six different weights on control confidence and box goal accuracy. High control confidence tools are curved to maximize the controllability of the box at the expense of accuracy of the box goal position.
• Figure 2: The schematic of the tool design pipeline. The designer designs and optimizes a tool mesh design. This mesh is used by the user to perform a task in a simulation environment using a fixed plan and a standard robot controller. The task variable in the simulation is tracked by the evaluator and the tool performance is evaluated. A confidence value is attributed to each tool and is used to compute the free energy objective that the designer uses to optimize the tool design.
• Figure 3: The realization of the optimization process of the tool design. a) The evolution of the best polynomial curves of the tool per iteration. b) The free energy for the best candidate in the population per iteration of the cmaes objective. c) The top view of the box movement trajectories from its start location (black circle) to the end location (black cross) while using the best tool per iteration. d) The final optimal tool design for $p^X=40$.
• Figure 4: The free energy objective rewards a decrease in goal error (before $t=1s$) and penalizes the decreases in tool's controllability on the box (after $t=1s$).
• Figure 5: The tools designed by the objectives with $p^X = 0,10,20,30,40,50$. The tool designed exclusively by optimizing control confidence ($p^X=0$) is highly curved and perfectly encloses the box for higher controllability, while the tool designed by squared error ($p^X=50$) is straighter and prioritize to push the box closer to the goal at the expense of controllability.