Sandhya Subramani: The Robot You Can Just Talk To Is Already Here — BigGo Finance

The most telling moment in Sandhya Subramani’s robotics demonstration came when she asked a small rover named Scout to “do something complex.” Scout replied, “I called this my signature performance,” then did nothing. A beat later, it turned off its headlights. Then it tried to summon a “rover speak” function. This was not a scripted failure — it was a live demonstration, on stage, of what happens when you hand a physical machine to an LLM and ask it to improvise.

Subramani, an AWS technologist, was speaking on the AI Engineer podcast, and her central claim is that the bottleneck in robotics is no longer hardware or even policy training. It’s orchestration. And the fix is surprisingly simple to describe, if radical in implication: stop programming robots. Start telling them what you want.

The Live Demo: A Rover You Can Talk To

Scout is not an industrial prototype. It’s a small rover retrofitted with a Raspberry Pi and a 4G SIM, physically located in New York City while Subramani controlled it remotely. During the episode, Scout’s observable behaviors included turning on headlights and greeting the audience, spinning 360 degrees on command, reporting what it saw, and answering an unprompted question about how many people were in the room. The rover counted two people near the stage — a real-time perception feat that hints at the system’s situational awareness.

What made the demo compelling was not any single capability. It was the lack of a script. At one point, Scout fell off the stage. It was instructed to back off and recover — and it did. When asked to analyze the room and identify “the best looking person,” Scout chose Subramani, citing her “confidence stride” and “khaki jeans and sneakers.” The reasoning was coherent, the action autonomous, and the outcome unplanned.

This unpredictability is not a flaw to be ironed out, Subramani argues. It’s the feature. “So now if we wanted to do more things than just the tasks it’s trained on,” she said, “give it an agent and see what it can do.”

The Architecture: Four Layers and Five Lines of Code

The technical backbone is Strands Agents, an open-source framework built by AWS. Subramani described a four-layer architecture that moves instructions down and observations up:

The framework supports more than 40 different robots across eight categories, all accessed through a single “robot tool” call. Getting started requires five lines of code — a detail Subramani emphasized as intentionally low-friction. Training happens on the cloud using AWS’s Agent Core, while inference and policy execution can happen on the edge for low-latency response. Strands decides at runtime whether to call the cloud or the edge, balancing training throughput against execution speed.

This hybrid design matters for economics. If a robot can lean on cloud inference for complex reasoning but fall back to edge execution for split-second physical decisions, the cost structure of autonomous robots starts to look more like a software service than a hardware deployment.

Three Brains, One Rover

Scout runs three Strands agents simultaneously, each with a distinct role. A thinker agent continuously assesses the environment and decides what to do next, even without explicit instruction. A communication agent connects Scout to Telegram and a web app, enabling remote conversation and command execution. A voice agent handles spoken interaction — disabled during the episode because it would have interrupted Subramani’s presentation.

This multi-agent design is what gives Scout its apparent autonomy. The thinker runs constantly, so the rover can proactively comment on the stage setup, notice people, and decide to move toward them. The communication agent means Subramani can chat with Scout from anywhere via Telegram when she’s not home. The result is a robot that behaves less like a tool and more like a presence — one that can be interrupted, redirected, and reasoned with.

The system prompt is the unsung hero here. It tells Scout what each of its pre-existing tools is for — headlights, movement, speech — and includes safety guardrails. The LLM reads those tool descriptions and matches them to the user’s intent. This is the mechanism by which the agent knows which tool to invoke, and it is the reason Scout can handle novel instructions without retraining.

What’s Under the Hood

Subramani displayed her configuration screen during the episode, revealing the specific components powering Scout’s brain:

The choice of Claude Opus 4.8 is notable. It places the reasoning burden on a frontier LLM, not on a specialized robot model. That’s the entire thesis in one configuration choice: general-purpose intelligence, pointed at physical hardware.

The Data Flywheel No One Planned

A secondary but significant insight is that Scout’s autonomous behavior generates training data as a byproduct. Subramani explained that she can manually move Scout to navigate in a desired direction, creating training episodes. She can observe how Scout reasons and responds to questions, then use that data to improve its performance.

This creates a feedback loop with compounding effects: