Here's how cloud agents are improving production robot performance today. (Side note: I love these @AmbiRobotics blog posts. Ambi is one of the few production robotics companies willing to go deep on its technical work in public.) One problem that Ambi noticed for their sorting robots is that some packages barely fit the bag opening. Deformable shapes or sensor noise can cause the robot to misjudge the package size, so it can place them in the wrong orientation. This means the package doesn't fit, the task fails, and the system has to retry. Ambi estimates that this costs about 31,000 sorts per year. Obviously, solving these failure modes is highly attractive; reducing 31,000 extra sorts creates significant value through faster throughput. Unfortunately, engineers still need weeks to analyze failures, propose fixes, and test them. Instead, if Ambi can plug AI coding agents into the loop, they can iterate much faster. So they built the AmbiOS Agentic Robotics Harness. The agents don't solve the problem by writing arbitrary code. Instead, they can edit a bounded computation graph composed of existing robot skills. This means the agents can move quickly while still keeping safety checks and error handling intact. The harness gives the agents two key pieces of infrastructure: 1) A cloud dataset of real production events, filterable by known success and failure modes 2) A simulator that replays those events with the agent's proposed changes So the agents pull failures → hypothesize → edit the skill graph → test in sim → ship to a production A/B test. In their experiment, these agents proposed three candidate fixes in just 10 hours. Of the three, one added 4.2 packages per hour, about 15k more sorts per robot per year, and now runs on 30% of Ambi's U.S. fleet. I love how simple yet powerful this harness is. I think every robotics company with robotics in production should strive to accomplish such a loop.

What does it take to build a robust harness for a robot foundation model? A foundation model takes in images or video, sometimes with language instructions, and outputs robot actions. The model is a black box, and you can't guarantee its output is safe to run on real hardware. That's why the harness is important. It conditions what goes into the model, then checks and sometimes modifies what comes out, so the robot can act safely, reliably, and quickly. In conversation with @jmahl42 on the latest Scaling Robotics episode, we walked through PRIME-1 by @AmbiRobotics , a foundation model for picking. PRIME-1 takes in images and outputs a pick action, roughly an end-effector pose. The harness takes it from there: → Builds the full motion plan from that pose and optimizes it for speed → Checks that the robot stays in bounds and won't collide with anything → Throws out actions that fail those checks → Knows when to stop and hand a problem to a person The human-in-the-loop makes the system robust to edge cases. When a pick fails, it's not always optimal to send the problem back to the model. Sometimes the harness needs logic to exit the AI loop and bring in a human. Models predict what to do, and the harness dictates what the robot actually does. Harnesses are how we will get foundation models from demos to production!

Here's how @AmbiRobotics thinks about deploying robots into existing warehouse systems. 1) Keep integrations simple A robot shouldn't need ten pieces of information from upstream systems at ten different moments just to function. Whatever touchpoints are put in place need to be small and well-defined. This allows a robot to unlock generality. 2) Rate matching Warehouses run at a cadence. Items get picked at a rate; boxes move at a rate. If your robot is too fast or too slow, the rest of the system is out of whack. Ambi often connects to sorters, the big conveyor systems that carry packages across a room and divert them down slides. If a burst comes in faster than the robot can handle, that's a problem for the entire line. To solve for this, they introduced buffering 3) Design around line conditions Ambi has a blue bin they cart around next to AmbiSort. This is sometimes odd to customers, who ask if they can just connect it to the conveyor. But at deployment, bins are more optimal. The bin decouples the robot from the line. They can put the robot anywhere in the building, fill a bunch of bins, queue the work, and run it when ready. This builds in the buffer that makes rate matching possible. -- In conversation with @jmahl42 in the latest episode of Scaling Robotics!

I've enjoyed getting to know Jeff because I love his balanced approach to building and deploying Physical AI systems. On one hand, he is a pragmatic technologist. @AmbiRobotics has deployed 100+ robots that do real work for some of the world's biggest logistics operations. On the other, he's a PhD from @UCBerkeley, and his team's academic roots mean they're constantly pushing the frontier of end-to-end models. The result is a harmony of both rapid deployments and rapid progress on Ambi's autonomy stack. More companies and leaders should operate like this, and I'd encourage you to check out this episode with Jeff, out now.

Another harness for a learned robotic policy in production. We've seen a lot of cool demos of these policies in the last 10 days. The thing that companies are trying to solve is how to deploy them in production. @AmbiRobotics CARGO is one such answer. CARGO is a planning model. It decides which box to stack next and where to place it, so the container fills as densely as possible. The picking and placing are handled by manipulation skills that already exist in Ambi's capabilities. CARGO is the director that tells the manipulation what to do. We've seen other companies pair a learned manipulation policy with classical controls as a safety harness. Ambi moves the learned policy up a layer. Here it does the planning, and the harness keeps that planning physically grounded. CARGO only gets to choose from placements that are already deemed "feasible", meaning they are collision-free, stable, and executable given the real limitations of the hardware. It can't propose a more optimal solution that the robot can't physically reach, or one that would topple the stack. Whether the learned policy is for manipulation, planning, or both, there is a clear trend of classical harnesses being used to ensure movements are safe and physically grounded. @ambi

This is an important, trailblazing collaboration between @pickle_robot and @ambirobotics that more teams should pay attention to: Any Physical AI company that has sold into logistics knows how hard it is to get systems onto a production warehouse floor. Autonomy and reliability are, of course, important, but the conditions around the existing warehouse are often the bigger obstacle: - Do you integrate with my WMS? - How will I have to rearchitect my warehouse to fit your system? - Do you play nicely with the systems I've already sunk a ton of money into? Clearing these objections takes huge effort that has nothing to do with your product, and it usually means fighting legacy systems and incumbents who don't want to interoperate. When two Physical AI companies pre-integrate, they reduce that friction for the customer and cumulatively solve a larger share of work. The buyer gets systems that interoperate out of the box, which makes "saying yes" a lot easier. Great work by both teams. I want to see more of this across the industry!