# General Robotics launches agents that assemble and debug robot deployments > Source: > Published: 2026-09-09 19:18:10+00:00 # General Robotics launches agents that assemble and debug robot deployments **Ashish Kapoor's GRID built a dual-arm lab workflow in four hours, then reused the fixes to deploy later skills in 10-15 minutes, General Robotics says.** By [RuntimeWire Staff](/author/runtimewire-staff) ยท Published Primary source: [General Robotics on X](https://x.com/genrobotics_ai/status/2097713940893679902) ## Why it matters Robot hardware and AI models are advancing faster than deployment. If GRID can reuse validated fixes across sites and machines, General Robotics can productize the scarce engineering work that keeps industrial robot projects stuck in pilots. [General Robotics](https://genrobo.ai/?ref=runtimewire) co-founder and CEO [Ashish Kapoor (@akapoor_av8r)](https://x.com/akapoor_av8r?ref=runtimewire) launched Auto-Engineering for GRID on September 9th, giving software agents responsibility for assembling and debugging much of the stack required to make a robot perform a new task. The launch, described in [a thread on X](https://x.com/genrobotics_ai/status/2097713940893679902?ref=runtimewire) and a detailed [technical post](https://www.generalrobotics.company/post/introducing-auto-engineering-for-robotics?ref=runtimewire), extends a line of work Kapoor pursued during 17 years at Microsoft. He led Microsoft's autonomous systems and robotics research group and created AirSim, an open-source simulator for autonomous vehicles and drones. His work there treated simulation as a way to test complex systems before sending them into the physical world. Auto-Engineering attempts to automate the engineer who builds and repairs that bridge. Kapoor founded the Redmond, Washington-based General Robotics in 2023 with [Sai Vemprala (@saihv)](https://x.com/saihv?ref=runtimewire), Dinesh Narayanan and Shuhang Chen. Vemprala and Chen also worked in Kapoor's Microsoft group. Vemprala, now General Robotics' CTO, had researched multimodal learning, computer vision, reinforcement learning and simulation at Microsoft after earning a robotics Ph.D. from Texas A&M University. Their shared thesis has remained consistent since General Robotics began under the name Scaled Foundations: useful robot intelligence will require reusable software components, common interfaces and simulation infrastructure across different machines. The founders are trying to avoid rebuilding a separate stack for every arm, humanoid, quadruped, wheeled robot and drone. ### The engineering layer Kapoor wants to productize A robot model can recognize objects or generate motions without knowing how a particular machine's joints, cameras, grippers, calibration, control loop and safety constraints fit together. That integration work still falls to specialists who connect hardware, build simulations, collect data, train policies, evaluate failures and repeat the process on the physical robot. "That last layer is missing," Kapoor told [GeekWire](https://www.geekwire.com/2026/general-robotics-led-by-microsoft-vets-says-its-ai-has-cut-robot-setup-from-a-month-to-hours/?ref=runtimewire). Auto-Engineering turns [GRID](https://www.generalrobotics.company/grid?ref=runtimewire), General Robotics' development and deployment platform, into an agentic system organized around four harnesses. One ingests a robot's physical configuration and control interfaces. A second constructs simulation environments and generates experience. A third chooses how to create a skill, including composing existing components, producing synthetic demonstrations, requesting human demonstrations or fine-tuning a vision-language-action model. The fourth deploys the result, evaluates it on hardware and traces failures back through the stack. The important design choice is what happens after a task works. GRID retains validated robot descriptions, calibration corrections, simulation components, perception models, training recipes and deployment repairs. A later task can reuse that work instead of beginning with an empty integration project. ### A lab demo exposes the difficult parts General Robotics demonstrated Auto-Engineering on a laboratory workflow using two Flexiv arms. The initial goal was to pick up a test tube, pass it between the arms and pour its contents into a beaker. GRID constructed a hybrid simulation that coupled a fluid solver written with NVIDIA Warp to MuJoCo's rigid-body physics. It then assembled object segmentation, grasp generation and collision-aware motion planning into a skill. The founders made the task progressively harder. When the beaker moved during pouring, GRID generated synthetic demonstrations and trained a reactive policy with behavior cloning and DAgger. Deployment exposed a perception pipeline that could not locate the moving beaker at the required control rate. GRID used output from the slower segmentation system as supervision for a lighter detector and tracker, then inserted that component into the deployed skill. For a stirring task that could not be expressed with the available motion primitives, Auto-Engineering requested human teleoperation data and fine-tuned a visuomotor policy. For swirling a flask, General Robotics supplied one phone video of a person performing the motion. GRID reconstructed the interaction in simulation and generated synthetic demonstrations without a robot demonstration. The sequence matters because it shows Auto-Engineering making different technical choices for different tasks. It also shows how quickly autonomy runs into ordinary engineering constraints. The system still needed human demonstrations for stirring, and real hardware revealed perception problems that simulation had not eliminated. General Robotics reports that the first working skill on the fresh Flexiv setup took about four hours, including roughly 20 minutes for robot ingestion and 10 minutes for the initial simulation. Later skills on the same setup took as little as 10-15 minutes because GRID retained the earlier integration and repairs. These are controlled demonstrations reported by General Robotics, rather than independent benchmarks across customer sites. ### Reuse is the commercial bet General Robotics is positioning reusable engineering knowledge as the layer between increasingly capable robot models and enterprises that still struggle to move beyond pilots. Robot manufacturers can supply the hardware, while model developers supply perception and control capabilities. GRID is meant to determine how those pieces should be combined, tested and repaired for a specific deployment. Kapoor told GeekWire that General Robotics has about 50 employees, roughly a dozen enterprise customers and revenue in the millions of dollars. Those operating figures come from Kapoor. HTX, the science and technology agency for Singapore's Ministry of Home Affairs, is one named customer. General Robotics has raised nearly $34 million, according to GeekWire. Its latest disclosed financing was an approximately $25 million April round led by [Construct Capital](https://www.constructcap.com/?ref=runtimewire), with participation from [Khosla Ventures](https://www.khoslaventures.com/?ref=runtimewire), [Accenture Ventures](https://www.accenture.com/in-en/about/ventures-index?ref=runtimewire), NVIDIA and [Valo Ventures](https://valoventures.org/?ref=runtimewire). Accenture confirmed its strategic investment and a partnership aimed at manufacturing, logistics and other asset-heavy industries, while keeping the investment amount private. The investor mix reflects the market General Robotics is pursuing. NVIDIA participated in the financing and its Warp software appeared in the demonstration. Accenture disclosed a strategic investment and a partnership focused on industrial sectors. The current demonstration establishes that GRID can preserve useful work across a controlled sequence of lab tasks. The commercial test is whether those gains hold across unfamiliar hardware, changing factory conditions and repeated production cycles. Kapoor has spent much of his career building simulations and tools for systems that must leave the lab. Auto-Engineering is his attempt to make the deployment expertise itself reusable software.