Abstract
Legged machines are becoming increasingly agile and adaptive but they have so far lacked the morphological diversity of legged animals, which have been rearranged and reshaped to fill millions of niches. Unlike their biological counterparts, legged machines have largely converged over the past decade to canonical quadrupedal and bipedal architectures that cannot be easily reconfigured to meet new tasks or recover from injury. Here we introduce autonomous modular legs: agile yet minimal, single-degree-of-freedom jointed links that can learn complex dynamic behaviors and may be freely attached to form multilegged machines at the meter scale. This enables rapid repair, redesign, and recombination of highly-dynamic modular agents that move quickly and acrobatically (non-quasistatically) through unstructured environments. Because each module is itself a complete agent, the bodies that contain them can sustain deep structural damage that would completely disable other legged robots. We also show how to encode the vast space of possible body configurations into a compact latent design space that can be efficiently explored, revealing a wide diversity of novel legged forms.
Introduction
In nature, animals exhibit an extraordinary range of body plans, each specialized for specific ecological niches. In contrast, the field of legged robotics has largely converged on a few fixed, canonical forms—primarily bipeds and quadrupeds. While these robots have achieved impressive levels of agility, their fixed architectures represent a significant limitation: they cannot easily adapt their physical structure to new tasks or recover from catastrophic structural failure.
Existing modular robotic systems, designed to address this lack of flexibility, often suffer from a "modularity tax." Because they are typically constructed from generic, low-power blocks, they are often limited to slow, "quasistatic" movements. This research introduces a high-performance modular system that bridges the gap between the flexibility of modularity and the athletic agility of specialized legged robots.
The Minimal Leg: A Complete Autonomous Agent
The foundation of this system is a single-degree-of-freedom (1-DoF) jointed module. Unlike traditional modular components, each leg is a "complete agent" equipped with its own high-torque motor, battery, and sensor suite.
A critical design feature is the 63.5° axis intersection, which allows the links to sweep a wide conic path. This geometry enables even a single isolated module to perform complex dynamic behaviors, such as jumping, rolling, and turning. By treating the module as the fundamental unit of both power and intelligence, the system eliminates the need for a central torso or a single point of failure.
Navigating the Morphological Design Space
Combining these modules into multi-legged robots presents a combinatorial challenge; there are billions of potential configurations. To navigate this vast space, the researchers utilized a generative design approach:
- Latent Compression: A Variational Autoencoder (VAE) was trained on 500,000 randomly generated configuration trees to compress the discrete space of body plans into a continuous 8D latent space.
- Optimization: Using Asynchronous Bayesian Optimization (BO), the system "fished" through this continuous space to identify the most efficient and capable designs for specific tasks.
- Simulation to Reality: Designs were optimized within the MuJoCo physics engine using domain randomization—varying parameters like friction and mass—to ensure that behaviors learned in simulation would transfer "zero-shot" to the physical world.
Amputation-Agnostic Control
To ensure resilience, the researchers developed a control architecture capable of handling radical structural changes. Rather than training a specific policy for every possible configuration, they utilized a Causal Transformer trained on sensorimotor data across a wide distribution of "damage scenarios."
This policy does not require explicit notification of damage. Instead, it learns to infer the current state of the robot's body—including missing limbs—solely through internal proprioceptive sensors. By treating a lost limb as a massive perturbation, the controller can adapt the robot's gait in real-time. This allows a robot to maintain locomotion even when reduced from five modules down to a single functioning unit.
Experimental Results
The system demonstrated remarkable performance across both individual modules and complex multi-legged assemblies:
- Individual Agility: A single module proved capable of rolling at 0.46 m/s and performing vertical jumps reaching 154% of its own length.
- Locomotion in the Wild: Optimized designs, such as a three-module "seal" and a five-module "lizard" with an actuated spine, were tested on diverse outdoor terrains including mud, tall grass, and tree roots. These robots maintained high-speed locomotion without any fine-tuning of their simulation-trained policies.
- Generalist Superiority: Interestingly, the amputation-agnostic generalist policy performed slightly better (105.3% speed) on an undamaged quadruped than a policy specialized specifically for that form, suggesting that training across diverse morphologies enhances overall robustness.
Discussion and Future Work
This modular approach demonstrates that high-performance agility does not require complex, multi-jointed limbs or fixed body plans. By decentralizing power and control, robots can become inherently more robust; a modular robot that is split in half effectively becomes two smaller, functional robots rather than one broken machine.
Future iterations of this work aim to address current limitations, such as the requirement for manual assembly. The integration of docking mechanisms could enable autonomous self-reconfiguration. Additionally, while the current system relies on proprioception, the addition of exteroceptive sensing (vision) would allow these diverse forms to navigate complex obstacles through proactive planning rather than purely reactive adaptation.
Summary updated