Task2Morph: Differentiable Task-inspired Framework for Contact-Aware Robot Design

TL;DR

This paper proposes a novel and general differentiable task-inspired framework for contact-aware robot design called Task2Morph, and embeds the mapping into a differentiable robot design process, where the gradient information is leveraged for both the mapping learning and the whole optimization.

Abstract

Optimizing the morphologies and the controllers that adapt to various tasks is a critical issue in the field of robot design, aka. embodied intelligence. Previous works typically model it as a joint optimization problem and use search-based methods to find the optimal solution in the morphology space. However, they ignore the implicit knowledge of task-to-morphology mapping which can directly inspire robot design. For example, flipping heavier boxes tends to require more muscular robot arms. This paper proposes a novel and general differentiable task-inspired framework for contact-aware robot design called Task2Morph. We abstract task features highly related to task performance and use them to build a task-to-morphology mapping. Further, we embed the mapping into a differentiable robot design process, where the gradient information is leveraged for both the mapping learning and the whole optimization. The experiments are conducted on three scenarios, and the results validate that Task2Morph outperforms DiffHand, which lacks a task-inspired morphology module, in terms of efficiency and effectiveness.

Figures (6)

• Figure 1: In the scenario of box flipping, the position and size of the box can be regarded as task features since they are highly correlated with task performance. For different boxes, the feature distribution is different, which affects the optimal morphology of the robot. We aim to abstract the related task features and use them to inspire robot design directly.
• Figure 2: Task2Morph Overview. It is mainly composed of two parts, the left part is task-inspired morphology generation which is the focus of our work. It consists of three parts: task abstraction, mapping from task features to morphology parameters and experiment backpropagation. The right part showcases the co-design framework known as DiffHand, which is used to fine-tune our initial morphology and optimize the controller.
• Figure 3: The average performance of the Task2Morph, Task2Morph-F, DiffHand and DiffHand-F in three scenarios. Figures report the mean and standard deviation of the loss for 20 randomly selected tasks in every scenario. Some areas in the subfigures are scaled to clarify the convergence speed of the algorithms. The horizontal axis of each plot stands for the number of simulation episodes, while the vertical is the task-specific objective loss $\mathcal{L}$. We smooth out the curves with a window size of 5.
• Figure 4: Experimental results of DiffHand and Task2Morph-F under three different tasks in Flip Box scenario. Each task comprises two subfigures, denoted as (a) and (b), illustrating the optimized final morphology for DiffHand and Task2Morph-F, respectively.
• Figure 5: Comparison of objective loss of initial morphologies generated by Task2Morph and DiffHand. Randomly select 100 tasks in Finger Reach, Flip Box, and Rotate Plank scenarios, respectively, and each task gets the point with the loss of DiffHand on the horizontal axis and the loss of Task2Morph on the vertical axis when the algorithm converges. 70%, 75% and 72% tasks of Task2Morph achieve better performance, respectively, and the average loss value is reduced by about 194.05, 536.01 and 311.17.