Physical AI OS: copper-rs

Teams ship with Copper

Teams ship with Copper

Our last blog post highlighted a core belief: intelligence for the physical world must come from the physical world. An adjacent belief that gets less airtime, but is paramount for physical AI’s success: The model is not the product, models evolve all the time, the ecosystem of HW and SW that orchestrates autonomy with models is the product. 

Maturing the hardware for reliability in the field, validating safety requirements and recovering from errors with minimal downtime is quintessential for deploying the physical AI stack in the target environment (public roads, warehouses, closed off construction sites). Deploying in the target environment is what feeds the learning velocity of AI. This is a multi-year effort for a single embodiment. Replicating its success for dozens of embodiments, each with a different target environment requires embodiment agnosticism to be a core design principle of both SW and HW.

Our Spatial Intelligence Kit, at its core follows an embodiment agnostic hardware design principle, coupled with copper-rs (an open source physical AI OS) we’ve optimized for high learning velocity.

100x better latency, 10x higher logging throughput

100x better latency, 10x higher logging throughput

Optimal learning velocity for physical AI

1. Model ≠ Moat. Viewing the physical AI stack as a network of interlinked engineering systems, it's clear that the entire stack is as strong as its weakest link. A release candidate AI that has shown improvements on offline benchmarks still has to go through HW and physical AI OS integration, validation across scenario variations, traced against safety requirements, as well as monitored in the field. The learning velocity for AI is set by the maturity of the entire physical HW and SW stack and not just by the training time of the models.

Copper helps accelerate Yaak by being statically described, having a deterministic runtime that is built in the open, and by having safety certification in mind from day one. The physical AI OS is designed for the model, not retrofitted around it.

2. Build loop ≠ AI loop. Quite often within physical AI, reflex is to treat learning velocity as an antagonist to safety. Testing each build release spans weeks or months and thus performance shortcomings age quietly, before the results flow back upstream. By bringing the physical AI stack within the model training loop means that a multimodal log replays deterministically with HIL (hardware in the loop), every change gets tested, and any regression on safety requirements surface when they are cheap. The velocity of the feedback loop is an enabler for safety.

Copper's build, run, and replay turns every recorded field-run into a repeatable regression test. Integrated with Nutron, regression gets isolated and surfaced instantly.

Deterministic replay of an end to end algorithm

Deterministic replay of an end to end algorithm

3. Demos ≠ Deployment. The deployment gap in physical AI is ever growing as product demos don’t hold water when shipped. A fleet generates data in the target environment, the data improves the model, the model is deployed back to the fleet, and every turn of the loop gets faster as the deployment gap is bridged. 

SIK with Copper is the physical AI stack, forward deployed: collecting multimodal data in the target environment, to tune the AI and actuate the physical asset on-edge for any embodiment.

We built the Spatial Intelligence Kit (SIK) and Nutron to collapse the long deployment timelines: 

1. Pair our Spatial Intelligence Kit with the asset

2. Operate the asset as usual as our AI learns from observing your operations

3. You get pinged when it's ready to be part of your fleet as an autonomous worker

We are deploying SIK paired with multiple embodiments on a million physical assets with Copper in the next years.

Yaak transforms physical assets into autonomous workers. If you’re running a fleet of pallet jacks, forklifts, wheel loaders, street sweepers and you’re ready to explore autonomous operations, let’s talk.