Before You Budget for Physical AI: What Agility's IPO Financials and Lab Research Actually Tell Enterprise Buyers

Enterprise buyers are being approached with humanoid robotics pilots at a moment when the public financial record and the academic research literature are both signalling the same thing: the technology is progressing, but the unit economics and capability gaps make near-term deployment a high-risk commitment for most operations. Understanding where the gaps actually sit, and why they exist, is more useful than a vendor roadmap.

Companion piece to our broader work on physical AI deployment readiness. See Physical AI Deployments: Why Robots Fail Where Software Succeeds for a technical breakdown of why physical AI systems break in production environments.

What Agility's IPO Numbers Actually Tell You

Agility Robotics is the first humanoid company to offer public financial disclosure, which makes its filings an unusually transparent window into the sector's commercial maturity. The headline figures are stark: approximately $1.8M in revenue against roughly $140M in operating losses. That ratio is not a startup anomaly to be explained away. It reflects the genuine cost structure of producing, deploying, and supporting humanoid systems at early scale.

The losses are not primarily marketing or overhead. They reflect the engineering depth required to keep bipedal systems operational in real warehouse environments, the support burden per unit, and the cost of iterating on hardware that cannot be patched with a software update. For enterprise buyers, the implication is direct: any vendor offering you a pilot at a fixed per-unit fee is either subsidising that cost or has not yet encountered the full support surface of a live deployment.

The revenue figure also tells you something about market absorption. After years of development and significant Amazon partnership activity, the commercial revenue base remains narrow. Demand is not the constraint. Operational readiness at the price points enterprises can justify is.

The Vision-Only Ceiling in Dexterous Manipulation

Most deployed and near-deployment robotic systems rely on vision as their primary sensing modality. For pick-and-place tasks involving well-separated, consistently oriented objects, vision is sufficient. The ceiling appears when tasks require contact reasoning: inserting a component, handling deformable materials, or maintaining grip under variable load.

Research is now quantifying that ceiling precisely. DexTacWAM, a visuo-tactile world-action model tested across six contact-rich manipulation tasks on a 22-DoF bimanual platform, achieved an average task score of 70.6 compared to 38.0 for the strongest vision-only baseline. Removing tactile world modelling from the same architecture dropped the four-task mean from 74.7 to 26.6, even when tactile sensor features remained available to the action model (Yuan et al., arXiv 2026). The performance gap is not marginal. It is the difference between a system that can handle contact dynamics and one that cannot.

The commercial translation is straightforward. If your target tasks involve contact-rich manipulation, vision-only systems will underperform at rates that make production deployment untenable. Tactile sensing hardware and the training pipelines to support it are not yet standard in commercially available humanoid platforms.

The Data Acquisition Bottleneck

Even where the sensing architecture is sufficient, policy training requires large volumes of demonstration data. Real-robot data collection is expensive, slow, and constrained to the physical environments where the robot operates. This is not a software problem that scales easily.

One active research direction attempts to close this gap by converting human video into robot-usable training data. The HuRo pipeline processes heterogeneous human activity videos into robotised episodes with retargeted actions, producing approximately 630,000 episodes from five human-video sources. Pretraining on increasing volumes of this data improved task completion from 51.5% to 80.3% and out-of-distribution completion from 34.9% to 72.2% across four real-world manipulation tasks (Jeong et al., HuggingFace 2026).

Those results are meaningful progress, but they also reveal the scale of data required to achieve generalisation. 630,000 synthetic episodes to support four tasks, with out-of-distribution performance still below 75%, indicates that policy generalisation to novel objects, layouts, or lighting conditions remains a genuine open problem. Enterprise environments are, by definition, full of the kinds of variations that stress out-of-distribution performance.

Capability Thresholds That Should Gate Pilot Commitments

Given the financial and technical picture, the question for enterprise buyers is not whether to engage with physical AI but which capability thresholds should be demonstrated before committing pilot budget. We would frame three concrete gates.

Task Constraint Specificity

The vendor should be able to define the exact object geometries, weight ranges, surface properties, and positional tolerances within which their system operates reliably. Any answer that references general adaptability without specific bounds is not ready for production evaluation.

Demonstrated Out-of-Distribution Handling

Pilots should include deliberate variation in object placement, lighting, and surface condition. Systems that perform well in controlled demonstrations but degrade under minor variation will not survive a real warehouse or production floor. Require quantified success rates across varied conditions, not curated demonstrations.

Support Cost Transparency

Ask for the fully loaded cost per operational hour, including maintenance, downtime, and remote support. Given Agility's disclosed cost structure, any figure that looks competitive with human labour at current wage rates should be interrogated carefully. The economics may improve, but buyers should not underwrite that improvement through their own pilot budgets.

Where the Research Pipeline Points for 2026 to 2028

The academic work points toward two near-term developments that enterprise buyers should track without over-indexing on. First, tactile sensing integration is moving from research platforms toward more practical form factors. DexTacWAM demonstrated that a pretrained vision model can be extended to tactile dynamics in approximately four hours of adaptation with around 100 demonstrations per task, without full tactile pretraining (Yuan et al., arXiv 2026). That is a meaningful reduction in the data cost of adding contact sensing.

Second, human-video robotisation pipelines are maturing as a way to reduce the cost of policy pretraining. The HuRo results suggest that diverse human-video data can serve as a foundation for generalisation, reducing the volume of expensive robot-collected demonstrations needed for task-specific finetuning (Jeong et al., HuggingFace 2026). Neither development makes 2026 the year to scale physical AI deployments. Both make 2027 or 2028 a more defensible window for structured pilots in constrained, well-defined task environments.

The appropriate posture for most enterprise buyers right now is active monitoring rather than committed deployment. That means engaging with vendors to understand their hardware roadmap and support model, identifying the two or three tasks in your operation where the capability thresholds are most likely to be met first, and building internal evaluation criteria before a vendor builds them for you.