Generating Freeform Endoskeletal Robots

TL;DR

This work tackles the limitations of purely rigid or purely soft robot design by introducing endoskeletal robots that combine internal joints with soft external tissues. It develops a voxel-based, massively parallel simulator to model soft-rigid interactions, a variational autoencoder–based latent design genome to encode diverse morphologies, and a universal graph-transformer controller trained jointly with evolution to optimize form and function. Through four diverse task environments, the authors show that an evolving latent space yields coherent, interpretable morphology genes and that morphology evolution substantially enhances locomotion performance beyond random design exploration. The framework provides a scalable, open-ended platform for benchmarking co-design methods in complex embodied systems and paves the way for future sim2real investigations and volumetric fabrication of endoskeletal robots.

Abstract

The automatic design of embodied agents (e.g. robots) has existed for 31 years and is experiencing a renaissance of interest in the literature. To date however, the field has remained narrowly focused on two kinds of anatomically simple robots: (1) fully rigid, jointed bodies; and (2) fully soft, jointless bodies. Here we bridge these two extremes with the open ended creation of terrestrial endoskeletal robots: deformable soft bodies that leverage jointed internal skeletons to move efficiently across land. Simultaneous de novo generation of external and internal structures is achieved by (i) modeling 3D endoskeletal body plans as integrated collections of elastic and rigid cells that directly attach to form soft tissues anchored to compound rigid bodies; (ii) encoding these discrete mechanical subsystems into a continuous yet coherent latent embedding; (iii) optimizing the sensorimotor coordination of each decoded design using model-free reinforcement learning; and (iv) navigating this smooth yet highly non-convex latent manifold using evolutionary strategies. This yields an endless stream of novel species of "higher robots" that, like all higher animals, harness the mechanical advantages of both elastic tissues and skeletal levers for terrestrial travel. It also provides a plug-and-play experimental platform for benchmarking evolutionary design and representation learning algorithms in complex hierarchical embodied systems.

Figures (18)

• Figure 1: An evolving population of endoskeletal soft robots were encoded by a low dimensional latent design genome with minimal morphological assumptions and optimized for locomotion across complex terrains in multiphysics simulation using a shared universal controller that was simultaneously learned alongside morphological design. Videos and code at https://endoskeletal.github.io.
• Figure 2: Designing endoskeletal robots. A population of 64 robot designs (A-C) was sampled from a multivariate normal distribution (D) centered about an initially random point in latent space. The behavior each design in the population was optimized in physics simulation for 20 epochs of reinforcement learning under a shared universal controller. The ten red behavioral traces to the right of A-C project the forward locomotion of these randomly generated designs (from 3D to 2D) into the right hand side of the page, at every other epoch of learning. The fitness of a design was based on the best performance it achieved during reinforcement learning. After learning, the mean and covariance matrix of the design distribution were then adapted by evolutionary strategies and a subsequent generation of 64 new designs was sampled from the shifted distribution (E). This process was repeated for dozens of generations (F-J) yielding an evolved design distribution (K) that encodes a population of designs (L-N) with much higher fitness (blue behavioral traces).
• Figure 3: A universal controller for freeform endoskeletal robots. The control of an endoskeletal robot (A) utilizes a graph data structure with edges that map bone connections within its skeleton, edge features that track the position, orientation and angle of each joint, and node features that track the position, orientation and movement of each bone (B). The controller also tracks proprioceptive and mechanosensory input from the position, velocity and strain of the robot's soft tissues (C), which is locally pooled into the center of mass of each bone (D) thereby transforming the soft tissue's sensory input from voxel space to a graph that aligns with the skeletal sensory graph (E). The combined rigid+soft sensory graph is then fed as input to a graph transformer (F). The graph transformer distils sensory signatures across the graph into updated node features (G), and updated edge features (H). The Actor (I) takes the updated edge features as input and outputs motor commands (J): the target rotation angle for each joint (K). The updated node features are pooled by globally by a channel-wise maximum across the node dimension (L) to retrieve a graph-level output (M), which the Critic (N) uses to predict a value function based on the robot's current state information.
• Figure 4: Simulating endoskeletal robots. Soft tissues (green masses) were modeled by a grid of Euler–Bernoulli beams (A) that may twist and stretch and directly attach to bones (blue masses; B and C) that follow rigid bodied dynamics with joint constraints (D). More details about the simulator can be found in Sect. \ref{['sec:sim']} and Appx. \ref{['appx:sim']}.
• Figure 5: Interpolating between three points in endoskeletal latent space. Designs sampled from a 2D slice of the learned latent embedding (A-O) and their internal jointed skeletons (A$^\prime$-O$^\prime$), where A$^\prime$ reveals the skeleton of A and cyan cylinders show the location of hinge joints between bones in each skeleton. The three corner designs (A,K,O) were drawn from arbitrarily selected latent coordinates and the 12 designs between them sit with equal spacing along a plane of linear interpolation. The visible smoothness, mechanical coherence, and geometric and topological expressiveness of the voxel-based latent space facilitated the co-design of morphology and control.