An SRI spinout is leveraging SRI technology to increase the dexterity and flexibility of our future robotic collaborators.
“Robotics is the ultimate embodiment of AI, bringing AI out of the 2D screen and into the 3D physical world,” says Vikrant Tomar. As the founder of Avsr AI, a robotics startup incubated by SRI’s ventures group and recently spun out by SRI, Tomar is aiming to chart the future of robotic intelligence.
Applying AI to robotics has long intrigued scientists. SRI developed Shakey, the first robot to navigate its environment through artificial intelligence, in the late 1960s. Recently, emerging generative AI (GenAI) models have brought a powerful new toolset into the robotics playbook. “One of the biggest impacts of GenAI, beyond chatbots, is multi-step long-horizon reasoning, which is critical for making robots autonomous,” observes Tomar.
To continue making robots more useful in a complex, dynamic world, researchers must overcome three fundamental challenges: perception, cognition, and dexterity. In other words, robots need to understand their environment, plan and reason about the different steps needed to accomplish a task, and be able to interact with or handle objects of different shapes, sizes, and rigidity. “The real challenge is creating a unified framework that solves these three problems together. Not just creating a robot planner that makes intelligent decisions,” Tomar explains, “but also an operational system that translates those decisions into real-time dexterity — especially when handling objects that require human-like sensitivity and control.”
The seeds of Avsr AI
Tomar’s journey into AI began in India and later took him to McGill University, where he pursued a PhD in deep learning and AI, with a focus on speech and language understanding.
Tomar began to explore Montreal’s nascent but growing startup scene. He spent ten months working at Nuance, a speech-focused SRI spin-out from the same group that created Siri. Deep learning and neural networks for speech recognition were beginning to show their true potential, and Tomar took advantage of the burgeoning AI space to found Fluent.ai, a pioneering offline, noise-robust speech recognition technology designed to support any language or accent and, in the process, landing commercial partnerships with companies like Xperi (the parent company of TiVo), Arm, and Bosch.
Avsr (which stands for “autonomous versatile social robotics”) brings together AI and robotics veterans, including previous colleagues from the Fluent.ai team, with a suite of groundbreaking AI and robotics technologies licensed from SRI that are positioned to accelerate the company into a prominent role in the future of robotics. The SRI relationship will allow a founding team with deep AI experience to build on top of IP like SRI’s XRGo robotic telemanipulation and telepresence platform as well as the Dexterous Manipulation (breakthrough low-compute spatial intelligence) and SUWAC (GenAI-based robotics planning technology from the Center for Vision Technologies in SRI’s Information and Computing Sciences division).
Real-world data for real-world performance
While the technology behind Avsr AI may be complex, the objective is simple: to safely deploy robots in real-world environments, fine tune their underlying AI models as performance data accumulates, and quickly get them to a point where they are delivering return on investment.
Autonomous vehicles are prime examples of intelligent machines that navigate the world around them in real time. It has taken massive amounts of recorded driving data to get to this stage, yet in the grand scheme of robotics, driving is relatively simple — it’s mostly a two-dimensional interaction. Getting robots to autonomously roam around the world in three dimensions and manipulate objects is a much more complex problem. That means we need more data, with a significantly greater level of detail. For many robotics use cases, that data simply does not exist, and the limited data that does exist is usually a result of robot training that takes place in simulated environments rather than real-world conditions.
Simulating real-world environments in labs and R&D warehouses provides valuable training data, Tomar explains, but these simulations have inherent limitations. Models trained purely in simulated environments typically struggle when deployed in real-world settings, with even the leading labs achieving less than 50% success on general tasks. Alternative approaches — learning from AR environments or human videos — are in early research. The gold standard remains real-world data combined with quick adaptation during actual deployment. As of now, effectively bridging the “sim-to-real” gap requires complementing simulation-based learning with real-world data and adaptation mechanisms.
“We’re focusing on verticals where robotic collaboration can drive significant and measurable improvements.” — Vikrant Tomar
As the Avsr AI team develops methods to collect high-fidelity real-world data with minimal effort, it is also advancing research on training systems to perform well with progressively less data. While real-world data is essential, reducing the effort required to collect it — and ultimately decreasing the amount of data needed for effective training — paves the way for scalable and cost-efficient robotic solutions. To this end, the Avsr AI team has developed a framework that integrates reinforcement learning with SRI’s award-winning XRGo system, enabling seamless data collection and swift development of dexterous skills.
For Avsr AI, the key to rapid return on investment lies in getting robots into situations where they can accumulate optimal real-world data, including edge-case scenarios that are hard to simulate in a lab — what Tomar calls an “early-to-world” approach. That means early deployment in actual workplace contexts, in a safe and controlled manner. Essentially, the goal is to develop functional and versatile robots capable of quick adaptation with only light human intervention while maintaining exceptional performance.
Where will robots go next?
Numerous industries are keeping a close eye on the future of robotics. Manufacturers expect robotics to solve persistent labor shortages. Energy, mining, and chemical companies are excited about the ability of robots to operate in field conditions that might be unsafe for human employees. Retailers see a potential for robotic platforms that can help them optimize the cost of restocking shelves and correct pricing and placement issues 24/7.
For Tomar, the possibilities are vast, but not unlimited. “We’re focusing on verticals where robotic collaboration can drive significant and measurable improvements, such as addressing labor shortages and enhancing safety in hazardous environments,” he says. “How eager are people to embrace robotic help or automation in the market?” As an example of a place where all stakeholders might benefit from autonomous and dexterous robots, he points to manufacturing and logistics. “These domains are ripe for robotic collaboration. The productivity and safety dividends here could be enormous for everyone, from business owners and workers to consumers,” Tomar suggests. He points to space exploration as another place where autonomous robots with advanced dexterity and manipulation skills have a distinct edge, given that the environment is naturally hostile to humans.
Regardless of the use case, Tomar expects, the results will be better if we can quickly get robots on the ground and (safely) experiment.
To learn more about Avsr AI or SRI’s ventures team, contact us.
