Agility Robotics has announced the development of a whole-body control foundation model for its bipedal robot, Digit. This model is designed to function as a motor cortex, enabling the robot to execute general-purpose tasks such as walking, lifting, and manipulating objects in a stable and reactive manner. According to the company, the model is built on a compact Long Short-Term Memory (LSTM) architecture and was trained entirely in simulation using NVIDIA’s Isaac Sim. The training process spanned several days of simulated time and involved exposing the model to a wide range of scenarios and physical interactions.
The use of Isaac Sim, NVIDIA’s physics-based robotics simulation environment, allowed Agility to scale the training process efficiently and to iterate rapidly without the constraints associated with real-world testing. By relying exclusively on simulated training, the company was able to generate diverse and high-volume data sets, including variations in terrain, object dynamics, and disturbances. This approach enabled the LSTM model to learn robust control policies that generalize beyond the specific conditions encountered during training.
At its core, the model is responsible for coordinating control signals across Digit’s entire system, including its limbs, torso, and actuators. The aim is to achieve whole-body coordination that is both reactive and safe, particularly when interacting with unpredictable elements in the environment. According to Agility, this control layer plays a central role in translating high-level behavioral commands into low-level motor outputs, effectively acting as a real-time mediator between decision-making and physical execution.
Agility Robotics positions this development as a foundational step toward broader autonomy for humanoid robots operating in logistics, manufacturing, and industrial environments. The company has previously emphasized the importance of general-purpose mobility and manipulation in human-centric workspaces. With this new control model, Digit is expected to handle a wider array of physical tasks, such as lifting heavy packages or navigating cluttered spaces, without the need for task-specific programming or extensive manual calibration.
In addition to supporting locomotion and object handling, the model is designed to enable adaptive behaviors in response to unexpected forces or environmental changes. For example, if Digit is nudged off balance or encounters a shifting load, the LSTM-based controller can generate appropriate compensatory movements to maintain stability. This level of reactivity is essential for robots operating in real-world settings, where physical conditions are often dynamic and not fully predictable.
Agility also emphasizes the potential of this architecture to scale across different tasks and environments. Because the model was trained in a domain-randomized simulation environment, it has been exposed to a broad spectrum of variations in sensor noise, friction, object mass, and surface conditions. This exposure is intended to enhance the controller’s ability to generalize and to reduce the need for retraining when Digit is deployed in new settings.
The company has not disclosed specific performance benchmarks or deployment timelines for the model, but it has indicated that this work will form the basis for future developments in autonomous robotic behavior. The foundation model is part of Agility’s ongoing research into control hierarchies, simulation-to-reality transfer, and scalable autonomy for robotic systems.
