Executive Brief — Physical AI: What to Do and What to Wait On¶
Last updated: 2026-07 · owner: Youngjin · volatility: medium ← to index
L0 TL;DR: Robotics is meeting foundation models, putting the industry at an inflection point. But what to invest in now differs from what to watch. This page draws that line without the hype — 5 minutes.
① Why now¶
Three verification signals have stacked up. First, Agility Digit has run for pay under a multi-year RaaS contract at GXO logistics sites, logging more than 65,000 hours as of 2025-11 with customer cross-confirmation — the best-verified paid humanoid work to date. Second, Figure 02 ran a verification pilot on a BMW line (whereas Figure 03's "8-hour autonomous shift" is only a CEO tweet with no independent verification). Third, open foundation models such as Physical Intelligence's π0 have been released under Apache-2.0, making commercial fine-tuning possible. That said, no independent third-party autonomy audit of humanoids exists anywhere yet (pillar-4), so large-scale deployment forecasts remain items to watch, grounded only in vendor and customer PR.
② What it means for our industry¶
Manufacturing: A path has opened to fine-tune open VLA models to your own tasks. Rather than training from scratch, the realistic approach is to teach a specific process with a small number of real demonstrations (pillar-2). Logistics: Locomotion-based (walking) robots are already deployed at paid commercial sites. Precision manipulation, by contrast, is still at the research/preview stage, so task scope must be kept narrow (pillar-4). Automotive: The training pipeline for moving policies trained at scale in simulation to real hardware has matured. Parallel simulation on cloud GPUs lowers the barrier to entry (pillar-3).
③ What is real and what is hype¶
| Verdict | Meaning | Representative areas |
|---|---|---|
| 🟢 Invest now | Verified foundational capabilities | Robot data pipelines · simulation infrastructure · synthetic data |
| 🟡 Soon (12–24 months) | Worth a pilot | VLA fine-tuning capability · edge inference stack · agent orchestration |
| ⚪ Not yet | Items to watch | Large-scale humanoid adoption · fully autonomous shifts |
This distinction is not arbitrary. 🟢 holds because data pipelines and synthetic data (pillar-1) and simulation infrastructure (pillar-3) are GA and usable in production. 🟡 reflects that VLA fine-tuning is GA but lacks public real-world cases (pillar-2), and that edge/agents (pillar-5) are early in verification. ⚪ comes from the verdict that large-scale humanoid adoption such as Optimus and Figure 03 remains at the demo/roadmap stage (radar). This verdict is not made once and left alone — a continuous verification system that scans every week (radar) updates it through promotions and demotions.
④ So what do we do first¶
The honest answer is "data first." The bottleneck in robot learning is not the model but the data, and the real-world recipe is almost always a three-stage mix of open-dataset pretraining → synthetic data augmentation → fine-tuning on a small number of real demonstrations (pillar-1). Set the order as three steps. (1) Stand up the data pipeline first to accumulate assets, (2) secure low-cost, high-diversity data with a simulation PoC, and (3) on top of that, verify a VLA fine-tuning pilot on a narrow task. LoRA fine-tuning is possible with a single GPU, so the pilot entry cost is low (pillar-2).
⑤ Why do it with AWS¶
Three verified facts are enough. First, there is an official AWS blog case that trained Unitree H1 humanoid RL on Isaac Lab + SageMaker HyperPod (noting explicitly that this is RL locomotion, not VLA) (pillar-2). Second, Amazon Bedrock AgentCore is GA with full support in the Seoul region, so you can put agent orchestration on it without data residency concerns (pillar-5). Third, you can self-host open models such as π0 (Apache-2.0) and OpenVLA (MIT) and use them freely without vendor lock-in (pillar-2).
If you want to start a review¶
1-day architecture workshop: We recommend a diagnosis of your data assets plus applying the judgment matrix above to your situation. Reach out via your AWS SA or GitHub.
Go deeper: Technical Guide · P1 Data · P3 Simulation · Decision Tree
owner: Youngjin · updated: 2026-07 · volatility: medium (the judgment matrix is updated on radar promotions/demotions)