Subramani frames the AI agent as an orchestration layer between a person and a robotAgent orchestration coordinates multiple AI agents, their tasks, dependencies, communication and shared resources toward a larger objective.. Instead of forcing users to learn device-specific commands, the agent interprets an ordinary request, selects the relevant tools and delegates the physical details to a robot policyA robot policy is the decision-making function that maps observations and goals to actions a robot should take in its environment. or vision-language-action modelA vision-language-action model connects visual observations and language instructions to actions that a robot can perform..
Her Scout rover demonstration uses three cooperating agentsA multi-agent system coordinates multiple AI agents that have separate roles, context, tools, or responsibilities. for perception and planning, communication and voice. The thinker agent inspects the scene and available capabilities, the communication agent maintains the interaction, and the voice agent lets a person speak naturallyA natural language interface lets a person operate or configure a system using ordinary written or spoken language. while the system turns that intent into bounded robot actions.
The architecture is deliberately hybrid. High-level reasoning can run in the cloud, while latency-sensitive control and safety remain closer to the robot. A shared tool package gives the agent a consistent interface across more than 40 robots in eight categories, separating what the robot should do from how a particular machine performs it.
Subramani also shows how the workflow produces useful training evidence. Each instruction, selected tool, observed result and correction can become an episode for evaluating or improving the underlying policy. The live demo includes execution failures, which reinforces her point that practical agentic robotics needs explicit guardrails, observability and recovery rather than an assumption of perfect model behavior.
Watch on YouTube
