Robotics Research
Embodied AI at the edge. Adaptive precision inference for robots that need to think and act in real-time, on constrained hardware, in industrial environments.
Modern robotics is converging with frontier AI through vision-language-action models - systems that can perceive, reason about, and act in the physical world. But these models are designed for data center compute budgets. A robot on a factory floor or in a home does not have a GPU cluster. It has watts, not kilowatts, and milliseconds, not seconds, to make decisions.
SARAL's robotics research attacks this gap directly. We develop methods for adaptive precision inference that allow large models to run on small hardware, embodiment-aware transfer learning that moves skills between different robot forms, and deployment architectures for edge robotics in industrial and domestic settings. This is an emerging research domain at SARAL, with active investigations and early-stage results.
Featured Publications
Adaptive Precision VLA Inference
Vision-language-action models are the foundation of modern robotic intelligence, but they are built for cloud-scale compute. This research develops adaptive precision techniques that allow VLA models to run on edge hardware - dynamically adjusting quantization levels based on task complexity and hardware constraints.
EAIT (Embodiment-Aware Intent Transfer)
Transferring learned behaviors between robotic embodiments with different morphologies, sensor suites, and actuator configurations. EAIT provides a framework for intent-level transfer that is agnostic to the physical form of the robot.
All Publications
Edge Robotics for Industrial and Domestic Applications
A survey and research agenda for deploying robotic systems on edge hardware across industrial inspection, domestic assistance, and infrastructure maintenance domains.
Active Research
Robotics is an emerging domain at SARAL. These are our active research directions with investigations underway and early-stage results being developed.
Industrial Inspection Robots
Developing edge inference pipelines for robotic systems performing visual inspection of infrastructure - transmission lines, solar panels, industrial equipment.
Domestic Assistance Prototyping
Early-stage research into adaptive robotic systems for domestic environments, focusing on intent understanding and safe physical interaction.
Multi-Modal Sensor Fusion
Combining vision, tactile, and proprioceptive inputs on edge hardware for real-time robotic decision-making in unstructured environments.