Why robot models must be edge models — and what that means for India
The LLM race is decided by GPU capital. The robotics race can't be — models have to run on the robot. That single constraint reshuffles the board, and it's why we believe competitive physical AI can be built from India.
Every serious AI company is competing in a race defined by compute: whoever buys the most GPUs trains the biggest model, and the biggest model mostly wins. That race is real, and India rents tickets to it — we don't win it.
But robotics is a different race, because of one physical fact: the model has to run on the robot. A manipulation policy that waits 400ms on a cloud round-trip is a policy that drops parts. A factory with unreliable connectivity cannot depend on an inference endpoint in another region. Latency, safety and cost all point the same direction: small, task-scoped models running on edge compute inside the machine.
The two races, side by side
In the LLM race, models are massive and server-scale, data is internet-scale text and video scraped once for everyone, and the moat is billions of dollars of training compute. In the edge robotics race, models are small and on-robot; data is embodiment-specific — simulation self-play plus field data from your own deployments; and the moat is hardware fit plus deployment data nobody else can generate.
That's not a consolation prize. It's a structurally different game where the inputs are ownership of hardware, a simulator of that hardware, and actual deployments — all three of which we have.
Why egocentric video can't get you there
A popular path to robot foundation models trains on human egocentric video. It's elegant, but it has a hard ceiling: human video never contains hardware-specific motion optimisation. A policy trained this way still needs adaptation layers bolted on top to work on your arm, with your actuators, your transmission compliance, your control rates. Somebody else collected that data, for other hardware, at their cost — and you inherit the adaptation problem.
Our alternative: task-level RL from self-play in simulation of our own arm. The policy is fully embodiment- and hardware-adapted by construction. The sim data is generated by us, at zero marginal cost. Recent work supports the bet — self-play RL beats human-demonstration initialisation for dexterous manipulation (Success Guided Sampling, CoRL 2026). The whole transfer problem reduces to a tiny sim2real component at the end.
The compounding loop
Here is the loop that runs our AI programme:
- Deployment: the robot runs on a factory floor; revenue and field data come in.
- Sim RL run: a task-specific policy learns from self-play in our simulation.
- Sim2real adapt: a small transfer step onto real hardware.
- Teacher models: the hardware-adapted task policies become compact teachers.
- Edge VLA: a mixture-of-tasks/experts (MoT/MoE) model finetuned from the teachers, with a frozen vision pipeline, running on-robot.
Each task policy slots in as an expert — tasks don't compete for model capacity. Every task our robots run feeds the loop, so the VLA fits our embodiment better over time and each new task gets cheaper to teach.
Why this is viable from India
- Capital-light by construction: edge models are small and task-scoped. No frontier-scale training runs, no GPU fleets — the compute bill is a rounding error on a deployment.
- Sim data at zero marginal cost: self-play generates unlimited, hardware-adapted data for the one embodiment we control completely.
- Field data as a by-product: the deployment data no one else has is generated by every task our robots run.
- Better fit, not more data: models trained on oceans of egocentric video need volume; ours need ownership — of the hardware, the sim, and the deployments.
The frontier labs will keep winning the frontier. But the robots that actually run in factories won't be powered by a 400B-parameter model in a datacenter — they'll be powered by small, verifiable, task-scoped policies that fit the machine they live in. Building those is a race decided by deployments. It's the race we picked, and it's a race India can win.
— The Piqoid AI team · our technology →