New transformer architecture enables robots to control unfamiliar bodies without retraining

A significant barrier to universal robot control has been the need to retrain AI systems whenever encountering a new robot body. A research team now demonstrates how a transformer-based architecture with shared modular recurrence can substantially improve an agent's ability to command unfamiliar robots without additional training.

The challenge of morphology diversity#

Creating a single controller that works across many robot designs remains computationally expensive and data-intensive. While recent work has attempted to build multi-robot controllers by encoding robot-specific properties as contextual information, this approach struggles when robots have dramatically different structures. The difficulty intensifies when agents must perform zero-shot generalization—controlling robots they've never encountered before—especially when those new robots have different dynamics, kinematics, or body topologies.

How modular recurrence works#

The researchers' key insight is that not all robot information needs to be fully specified upfront. Instead, they argue that incomplete contextual features can be recovered through modular interactions between different parts of the control system. Their solution employs a transformer-based architecture that leverages shared modular recurrence, allowing the network to learn how robot components interact and infer missing information in a structured way. This approach exploits the underlying modularity present in most robot designs without requiring exhaustive pre-specification of every robotic feature.

Testing across diverse robots#

The team evaluated their system on a large set of MuJoCo-based robots across four distinct environments. The results show substantial gains in zero-shot generalization compared to prior methods, with the agent successfully controlling robots featuring previously unseen dynamics, kinematics, and body topologies. This suggests the architecture can discover and reason about robot structure on the fly, enabling more flexible and efficient multi-robot control.