field notes

Physical AI
Training Data

more articles

  1. egocentric video

    RekaDaily-10k: 10,000+ Hours of Household Video From the Claru Network

    Reka released RekaDaily-10k, an Apache 2.0 egocentric household dataset collected through the Claru network: 10,865 raw hours and 10,200 processed, captioned hours across 6.37M clips.

    Read→
  2. humanoid robots

    Japan Airlines Humanoid Robots Haneda Airport 2026

    Japan Airlines plans to deploy humanoid robots for ground handling at Haneda Airport by 2026, creating the most demanding real-world test yet for sim-to-real transfer in contact-rich outdoor manipulation.

    Read→
  3. physical-ai

    Bezos Project Prometheus $10B Physical AI Infrastructure 2026

    Jeff Bezos's reported $10B Project Prometheus initiative targets the infrastructure layers — data, simulation, evaluation — that physical AI and robotics foundation models still lack, signaling a platform play valued at levels comparable to GPT-4's total training investment.

    Read→
  4. physical-ai

    π₀.₇ Foundation Model: Steerable Emergent Robot Capabilities 2026

    Physical Intelligence's π₀.₇ achieves 82.1% success on trained tasks and 47.3% zero-shot generalization across seven robot embodiments and 50+ manipulation tasks, according to the team's technical report (arXiv:2604.15483)—redefining what a single 7B-parameter generalist robotic foundation model can do without task-specific fine-tuning.

    Read→
  5. humanoid robots

    Humanoid Robot Training Data Requirements in 2026

    Figure AI, 1X Technologies, and Agility Robotics all depend on multimodal training pipelines where a single misaligned sensor timestamp can break sim-to-real transfer — here are the actual specs, sync tolerances, and annotation schema decisions that determine whether humanoid robot training data produces working policies or silent failures.

    Read→
  6. physical-ai

    Physical AI Training Data Provider: 2026 Decision Framework

    He et al. (arXiv:2510.21391v1) show that VLA policies trained on real manipulation data outperform sim-only baselines by 30–60% on contact-rich tasks — this framework helps ML engineers decide when to buy real-world physical AI training data versus generate synthetic.

    Read→
  7. physical AI

    Physical AI Training Data Guide 2026

    Google DeepMind's RT-2 required 130K real-world robot episodes to generalize across 700+ manipulation instructions — this guide breaks down exact data specs, collection pipelines, and quality criteria by robot type.

    Read→
  8. diffusion-policy

    Diffusion Policy Robotics: Training Data Specs 2026

    Chi et al.'s Diffusion Policy achieves 85.7% average success on Push-T with roughly 200 demonstrations (arXiv:2305.12171), but generalizing across objects, lighting, and embodiments demands 10–50× more data with specific diversity constraints that most teams underestimate.

    Read→
  9. training data

    Training Data for Robotics: The Full Pipeline in 2026

    Google DeepMind's RT-2 needed 130K+ real-world episodes before language-conditioned manipulation worked reliably—here is the spec-level pipeline that makes datasets like that possible.

    Read→
  10. robotics

    Best Robotics Data Providers (2026)

    Nine robotics data providers compared on capture modality, annotation, delivery format and engagement model — Claru, Mecka, XDOF, Human Archive, iMerit, Scale AI, Appen, TELUS Digital and Surge AI.

    Read→
  11. VLA

    Best VLA Training Data Providers (2026)

    VLA training data providers ranked on action labels, embodiment coverage and delivery formats — with an honest first screen on what your pipeline actually ingests.

    Read→
  12. training-data

    Gig Workers Training Humanoid Robots: Why Data Quality Beats Volume in 2026

    1X Technologies and Prosper Robotics have deployed hundreds of gig workers to collect teleop data at home, but the volume-first approach has a quality ceiling that determines whether humanoid policies actually generalize.

    Read→
  13. VLA

    VLM vs VLA: What's the Actual Difference? (2026)

    VLMs generate text; VLAs generate motor commands. Here's exactly where the architectures diverge, what training data each needs, and why the distinction matters for robotics teams.

    Read→
  14. VLA

    How Much Training Data Does a VLA Model Need? (2026)

    OpenVLA pre-trained on 970K trajectories fine-tunes in ~1.5 hours with 50–200 demos for simple tasks. Here are the concrete numbers for VLA data requirements across task complexity.

    Read→
  15. sim-to-real

    The Sim-to-Real Gap Explained: Why It Happens and How to Close It (2026)

    Four specific causes of the sim-to-real gap — visual domain gap, physics approximation error, sensor noise mismatch, and long-tail scenario absence — and what real-world data addresses each.

    Read→
  16. physical AI

    The Physical AI Stack: From Raw Sensor Data to Robot Action (2026)

    Layer-by-layer breakdown of how physical AI robots learn: perception (Depth Anything V2, ViTPose, SAM3), world modeling, policy learning (Diffusion Policy, ACT, π0), and language grounding.

    Read→
  17. egocentric video

    Best Egocentric Data Providers (2026)

    Egocentric data suppliers compared on capture, licensing, annotation and delivery — Claru, Mecka, Human Archive, Luel, Encord, Labelbox, Appen, Scale AI and Ego4D.

    Read→
  18. data enrichment

    Data Enrichment Pipeline for Physical AI (2026)

    How Claru's enrichment pipeline adds depth maps, pose estimation, semantic segmentation, and action labels to raw video to produce training-ready physical AI datasets.

    Read→