Physical AI

Study sandbox · last verified 2026-05-08

Physical AI in 2026 — research, projects, and a learning roadmap.

Ten cross-cutting research docs and eighteen hands-on projects, organized around the four-loop data flywheel: collect → curate → label → eval.

Start with the overview →Browse projectsSee the connections graph
10
Research docs
~635 min total reading
18
Hands-on projects
From laptop to H100
8
Phases
Data fluency → Strategy
4
Loops
Collect · Curate · Label · Eval

Research

All 10 docs →
00~30m
Overview — Physical AI in 2026

Cross-cutting synthesis of the field: data is the bottleneck, the data engine is the product, the modular/E2E pendulum is dissolving.

01~90m
AV industry and data

Who collects what at what scale: Tesla, Waymo, Mobileye, Wayve, Waabi. Fleet logs vs customer-shadow vs sim vs world models.

02~60m
Robotics foundation models

The VLA recipe — RT-X, OpenVLA, π0/π0.5, Helix, Gemini Robotics. Open X-Embodiment, demo collection economics.

03~90m
Simulation and synthetic data

The competitor map. Applied Intuition Simian, Nvidia Cosmos + Isaac, CARLA, Waymax, the death of pure-rendering shops.

04~120m
Labeling and data curation

The most important doc for the role. Data-engine philosophy, FiftyOne/SAM2/OpenSCENARIO tools, foundation models as labelers.

05~60m
World models and generative

Cosmos, GAIA, GR00T-Dreams, Wayve PRISM-1. World models as data engine and as eval substrate.

The 8-phase project arc

All 18 projects →
  1. Phase A
    Data fluency
    2 projects
  2. Phase B
    Labeling fundamentals
    3 projects
  3. Phase C
    Production hygiene
    1 project
  4. Phase D
    Simulation and world models
    4 projects
  5. Phase E
    Robotics adjacency
    2 projects
  6. Phase F
    Behavior, sim agents, closed-loop
    3 projects
  7. Phase G
    Active learning + capstone
    2 projects
  8. Phase H
    Strategy
    1 project

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.

01FiftyOne scenario mining→
02MCAP / ROS 2 / Foxglove plumbing→
03SAM 2 + Grounding DINO auto-labeling→
04BEVFormer 3D detection→
06Privacy + provenance pipeline→
07CARLA OpenSCENARIO scenarios→
15Bench2Drive closed-loop eval→
16Active learning loop→
17BDD100K mini data-engine (capstone)→
18Strategy 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.