You'll be the reason Anvil's data collection system goes from "the parts work" to "the whole thing works." Anvil is building the platform layer for Physical AI: robotics hardware and software that is radically more accessible than legacy industrial solutions. We own and operate our own manufacturing, and in our first 12 months we built and shipped 200+ robots to customers in 60+ countries: NVIDIA's GEAR lab (the team behind GR00T), Physical Intelligence (the frontier lab behind pi0), Qualcomm, Toyota, Google, and the robot learning labs at Stanford, MIT, Columbia, CMU, and Georgia Tech. The researchers who define this field run their experiments on our hardware, and many of them bought it with their own money. Devkits are the wedge, not the business. That volume has earned us a custom force sensing actuator partnership with major actuator OEMs, and the install base becomes both the distribution channel for our next generation robots and a platform for the data and software layers above them. The handheld data collection system is the front end of that strategy: the instrument that turns real world tasks into training data at scale. Our handheld data collection system lets people capture high quality demonstration data for training robots to do real tasks, wherever the task actually happens, with no teleoperation rig and no motion capture stage required. The situation you're walking into Anvil already has specialist engineers deep in each piece of this system: pose tracking, model training, SoC/chip integration, and industrial design. Each piece works, but nobody currently owns making them work together as one product a customer can actually pick up and use. That is the gap you would fill. Responsibilities Full technical ownership of pulling the handheld data collection software system together end to end. Acting as the systems and product engineer across specialist teams, including pose tracking, model training, SoC/chip, and industrial design. You will not personally do their deep technical work, but you are the one who understands how it all needs to fit together, catches the gaps between teams, and makes the call on what integrates now versus later. Building real infrastructure for parts of the system that do not have it yet, including turning the model team's scripts into something repeatable. What the first 100 days look like By day 30: ramped on every major piece of the system, including pose tracking, app software, cloud pipeline, and the model team's current workflow, well enough to know where the real gaps are, not just the documented ones. Has personally used the device to collect data end to end at least once. By day 60: has shipped a first concrete integration or reliability improvement that makes two previously disconnected pieces of the system actually work together better. By day 100: is the person other engineers, including those in pose tracking, model training, SoC/chip, and industrial design, go to when something is not fitting together. Trusted with real technical judgment calls across team boundaries without formal authority over any of those teams. Requirements You have done systems or product integration work for a robotic or perception based product before. You have been the person who made disparate technical pieces (hardware, perception, software, sometimes ML) work together into something a real user relied on. This matters more than depth in any single domain. You have genuine, provable depth in at least one of the following, along with familiarity with several of the others: Perception/robotics systems (SLAM, VIO, sensor fusion, mapping) Embedded application development: you have gotten real applications built, deployed, and running reliably on constrained or embedded hardware Cloud scale pose estimation or perception pipeline engineering: building systems that turn sensor data into trustworthy estimates reliably at scale ML workflow, training, and deployment engineering Physical AI and robot learning deployment, for example VLA style policies, getting a trained model to actually make a robot do something You can operate as a team of one across teams you don't manage, building enough technical credibility in each domain to know when something's actually wrong versus just unfamiliar to you. You default to shipping the smallest thing that actually closes an integration gap, rather than hacking something bespoke every time or overbuilding a general platform nobody's asked for yet. Based in or willing to relocate to Taiwan, the USA, or Trinidad & Tobago. This is the same role and scope regardless of location. Requirements Master's degree in robotics, computer science, electrical engineering, or a related field — or a bachelor's with equivalent hands-on depth. A PhD is not required; this is a product integration role, not a research role. Years matter less to us than trajectory. The typical shape is 3–5 years of engineering experience, but 2–3 unusually fast-growing years count fully if they were spent working directly alongside a staff or principal engineer on a shipped robotics or perception product. Our ideal candidate has been the second chair on a real integration effort — watched how the cross-team calls get made, made a growing share of them under a senior engineer's cover — and is ready to own the whole thing for the first time. What this role is not Not a role where you'll personally develop the core SLAM algorithms, train the models, or design the chip. Those are owned by dedicated specialist engineers and teams. Your job is integration and systems and product judgment across their work, not doing their jobs for them. Not a management role. No direct reports today. You will drive alignment and technical decisions across teams you have no formal authority over. The "tech lead" part of this role is earned through judgment and trust, not a title. Not a narrow software engineering role focused on one clean layer of the stack. If you want to specialize in a single domain and hand off everything else, this isn't the right fit. Not a role with mature ML and data infrastructure to inherit. Where that does not exist yet, you are building it, not maintaining someone else's.
