Reconfigurable legged metamachines that run on autonomous modular legs

Chen Yu, David Matthews, Jingxian Wang, Jing Gu, Douglas Blackiston, Michael Rubenstein, Sam Kriegman

Introduction

Traditional legged robots are typically designed as monolithic systems with fixed body plans. While these designs can achieve high performance, they lack the morphological flexibility seen in biological evolution and are often vulnerable to single points of failure. If a central "torso" or power source is damaged, the entire machine becomes incapacitated. This research introduces a paradigm shift by developing autonomous, single-degree-of-freedom (1-DOF) modular "legs" that serve as the fundamental units of both movement and intelligence. By combining these modules, researchers can create billions of possible "metamachine" configurations that are agile, decentralized, and inherently resilient to structural damage.

Hardware: The Leg as an Independent Agent

The core of this system is a minimal legged module consisting of two 3D-printed links connected by a single high-torque actuated joint. Unlike previous modular robots that rely on a central controller, each of these modules is a complete, autonomous agent. Every unit contains its own 6S LiPo battery, ESP32-S3 microcontroller, inertial measurement unit (IMU), and motor controller.

Despite the simplicity of a single degree of freedom, the module's geometry—specifically a 63.5° axis of rotation—allows for a complex "conic path." This specific design enables individual modules to perform diverse maneuvers, including jumping, rolling, and turning, which provides the foundation for sophisticated collective locomotion when modules are combined.

Morphological Search in a Latent Design Space

With up to five modules, the number of possible physical arrangements reaches into the hundreds of billions. To navigate this combinatorial explosion, the researchers utilized a Variational Autoencoder (VAE) trained on a dataset of 500,000 randomly generated configurations. The VAE compresses this discrete space into a continuous, eight-dimensional latent space.

To find the most effective designs within this space, the team employed Asynchronous Bayesian Optimization (BO). This allowed for the efficient sampling of the 8D latent space to identify morphologies in simulation that maximized locomotive efficiency and speed. This co-optimization of form and function ensures that the resulting metamachines are physically capable of high-performance movement before a control policy is even applied.

Amputation-Agnostic Control

To manage the diverse and potentially changing shapes of these robots, the researchers developed a "Causal Transformer" trained via Deep Reinforcement Learning. The goal was to create a control policy that is "amputation-agnostic"—capable of adapting to the loss of limbs in real-time without explicit knowledge of the damage.

By distilling expert policies from various damage scenarios into a single model, the robot learns to adjust its gait based solely on its internal sensory history. This approach allows the metamachine to treat a broken or missing limb as a change in dynamics rather than a system-wide failure, enabling continuous operation in unpredictable environments.

Experimental Results and Resilience

The system demonstrated remarkable performance across both simulation and real-world testing:

  • Individual Agility: A single module proved capable of jumping 37 cm (154% of its link length) and rolling at speeds of 0.46 m/s with a highly efficient Cost of Transport (CoT) of 0.26.
  • Collective Speed: Multi-module assemblies reached speeds of up to 0.63 m/s. Crucially, the rank-order of performance was preserved from simulation to reality, validating the optimization pipeline.
  • Extreme Robustness: The amputation-agnostic policy allowed a quadruped configuration to maintain 105.3% of its original speed even after sustaining damage. The robots remained functional even when modules were completely removed or became "dead weight" (attached but non-functional).
  • Zero-Shot Transfer: The robots successfully navigated unstructured outdoor terrains—including sand, mud, gravel, and deep leaf litter—without any fine-tuning on physical hardware. This suggests that the combination of modularity and domain randomization in simulation creates highly generalizable locomotion.

Discussion and Future Work

While the current system demonstrates unprecedented resilience, it still requires manual intervention for physical reconfiguration and repair. Future iterations could explore autonomous docking mechanisms to allow the robots to self-assemble or shed damaged parts. Additionally, while the VAE provides a powerful tool for design, the "reconstruction loss" inherent in compression means some highly specific configurations may not be perfectly represented in the latent space.

This work paves the way for a future of "Lego-like" robotics, where standardized, mass-produced modular units can be rapidly deployed to form specialized machines. By shifting the focus from fixed designs to adaptive, decentralized modules, this research provides a blueprint for robots that can survive and thrive in the most challenging environments on Earth and beyond.

Summary updated

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.
BibTeX
@misc{yu2025reconfigurable,
  author = {Chen Yu and David Matthews and Jingxian Wang and Jing Gu and Douglas Blackiston and Michael Rubenstein and Sam Kriegman},
  title = {{Reconfigurable legged metamachines that run on autonomous modular legs}},
  year = {2025},
  eprint = {2505.00784},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.00784},
}
Figure