Physical AI in 2026 — research, projects, and a learning roadmap.
Sixteen cross-cutting research docs (including a glossary, a canonical reading list, and a four-part automotive-industry history), three interactive tutorials, and twenty hands-on projects, organized around the four-loop data flywheel: collect → curate → label → eval.
Research
All 16 docs →Cross-cutting synthesis of the field: data is the bottleneck, the data engine is the product, the modular/E2E pendulum is dissolving.
Who collects what at what scale: Tesla, Waymo, Mobileye, Wayve, Waabi. Fleet logs vs customer-shadow vs sim vs world models.
The VLA recipe — RT-X, OpenVLA, π0/π0.5, Helix, Gemini Robotics. Open X-Embodiment, demo collection economics.
The competitor map. Applied Intuition Simian, Nvidia Cosmos + Isaac, CARLA, Waymax, the death of pure-rendering shops.
The most important doc for the role. Data-engine philosophy, FiftyOne/SAM2/OpenSCENARIO tools, foundation models as labelers.
Cosmos, GAIA, GR00T-Dreams, Wayve PRISM-1. World models as data engine and as eval substrate.
Interactive tutorials
All 3 tutorials →Self-contained guides with live simulations — open one full-screen and poke at it.
A five-part interactive series walking up the stack from the physical sensor to the data engine everything depends on — point clouds, perception, SLAM, corruption, robustness, and building a simulator. Grounded in published benchmarks; figures current as of early 2026.
Three interactive field guides on how self-driving cars know where they are: a tour of the SLAM back-end (ORB-SLAM, LOAM, Cartographer), the localization spectrum the industry is converging on, and why one LiDAR sweep does localization and perception at once.
A deep dive into 30 companies across self-driving and robotics in the US, China, and beyond — how they collect real-world data, manufacture synthetic data, auto-label, and evaluate. The 2025–26 shift: the flywheel collapsing into a single foundation/world model that drives, simulates, and evaluates at once.
The 8-phase project arc
All 20 projects →- Phase AData fluency2 projects
- Phase BLabeling fundamentals3 projects
- Phase CProduction hygiene1 project
- Phase DSimulation and world models4 projects
- Phase ERobotics adjacency2 projects
- Phase FBehavior, sim agents, closed-loop3 projects
- Phase GActive learning + capstone2 projects
- Phase HStrategy1 project
- Phase IFrontier extensions (Round 3)2 projects
Compressed-time critical chain
If you must compress 18 weeks into 6, this is the minimum-viable order. One thing not to skip: project 18 — the strategy memo.
Four loops
Every project is tagged by which part of the data flywheel it touches.
- COLLECT
Fleet logs, customer-shadow, simulation, world-model generation. Weeks–months.
- CURATE
Triage, embedding mining, scenario taxonomy, dedup, slicing. Hours–days.
- LABEL
Auto-label, human verify, distill, pretrain, fine-tune. Days–weeks.
- EVAL
Open-loop, closed-loop sim, scenario coverage, safety case. Continuous.