Robots Are Ready. Why Aren't They Working?

Robotics technology has surged ahead, yet its real-world application lags. This isn't a research problem; it's a people problem. For decades, the field has cultivated a narrow profile of contributors, leaving it unprepared for the very deployments now within reach. To truly see intelligent robots augmenting human labor at scale, robotics requires fewer per capita roboticists and more operators, reliability experts, and outsiders, shifting focus from a research subfield to a robust industry. This essay centers on intelligent robotic manipulators that learn from data, where the gap between research promise and deployment reality is currently widest.

The moment for widespread deployment is now. While roboticists have always aimed for real-world application, the necessary tools are finally crossing critical thresholds. Advances in pre-trained models, coupled with emerging techniques like behavior cloning and DAgger, offer a clearer path to success. Vision-Language-Action models are beginning to generalize, and a new wave of affordable, capable hardware is making systems economically viable. Robots still face challenges in cycle time and memory, but the fundamental blocker is no longer that "nothing works." Instead, the hurdle is building systems that prioritize reliability, integration, iteration speed, and unit economics, a different kind of engineering problem demanding different builders.

The Missing Application Layer

The intelligence stack in robotics is rapidly advancing, but a crucial layer connecting capabilities to economic impact is missing. In software, foundation models unlocked value when companies built application layers on top, creating integrated products and iterating in production. Robotics needs a similar layer, encompassing custom hardware, robust telemetry, purpose-built sensors, teleoperation infrastructure, fleet management, and safety systems. Waiting for perfect autonomy before building these systems is backward; they are what make autonomy economically meaningful. Labor, not necessarily full autonomy, is the immediate product. Reliable robotic labor, even with teleoperated fallbacks, creates value today.

Reframing the problem this way enables companies to deploy now, learn customer workflows, develop necessary hardware and software, and build operational expertise. The competitive advantage lies in customer relationships, domain knowledge, and integration depth, not just model performance. This application layer gap is fundamentally a talent gap, requiring operators, reliability-focused engineers, and product builders who can translate real-world constraints into shipping systems, precisely the skills historically undervalued by the field.

The Talent Bottleneck

The field of robotics faces a gatekeeping problem it can no longer afford. The rapid pace of innovation means new techniques are months old, and practical knowledge often resides in hands-on experimentation rather than textbooks. Deployment demands applied judgment in data curation, failure mode debugging, and iteration toward reliability, skills not exclusive to those with graduate degrees.

Despite surging interest, hiring remains narrowly concentrated in a few large humanoid robot companies. Outside these centers, hiring is conservative and credential-heavy, locking out many operators, software engineers, and product builders from adjacent fields. Graduate degrees have become de facto requirements, and experienced roboticists are often treated as scarce specialists rather than force multipliers. This slows execution and limits entry points when more builders are critically needed. While communities like Hugging Face lower entry barriers, hiring practices lag behind.

Robotics involves unique hardware and software challenges, but difficulty alone doesn't justify insulation. Fast-moving fields like aerospace routinely grant early-career engineers significant system responsibility. Cross-pollination is essential for growth; a resistance to outsiders guarantees a talent bottleneck. Investing in early-career talent and candidates from adjacent fields, prioritizing learning potential over existing credentials, is vital.

Deployment as the Forcing Function

Deployments must be treated not as an endpoint but as a crucial driver for the next cycle of research. Real-world use exposes failure modes, reliability constraints, operator workflows, integration challenges, and cost pressures that lab environments miss. These insights should inform every layer of the stack, from hardware design to model training. Waiting for perfect, universal robots is an excuse for inaction; the goal is a robot that delivers value today, even if narrowly constrained and heavily teleoperated.

The scaffolding for early deployments, safety systems, monitoring, remote intervention, and operational infrastructure, forms the foundation for future autonomy. Deployment expands the system's robustness, revealing bottlenecks in software, data pipelines, and hardware, and highlighting mismatches with operator workflows. Each deployment generates feedback that refines models, software, and hardware design.

Companies closest to deployment are best positioned to capitalize on future model improvements. Legal AI platform Harvey, for example, built relationships and domain expertise within law firms using early, imperfect models, creating a distribution moat that outpaced competitors. As co-founder Gabe Pereyra noted, "Don't build for the current capabilities of models today, build for where the models are going to be." Robotics companies focused on operationalization, integration, and deployment infrastructure now will compound on every subsequent model improvement.

Researchers deserve credit for advancing the field, but a summative transformation is needed. We require a new wave of builders focused on deployment, iteration, and integration to solve the operationalization challenge. Deploying robots into real businesses is a first-class engineering problem, distinct from research, demanding its own dedicated talent. Robotics needs fewer roboticists per capita, but more operators, designers, infrastructure engineers, salespeople, support staff, and young talent willing to own deployments end-to-end. It must become more open and more grounded in the real world, with value creation as the ultimate benchmark. Robots only change the world when they leave the lab. It’s time to build. This article was originally published on the a16z Blog.