BLO LAB (Blomega) builds wearable sensing hardware, the GX-1 glove and HC-1 head camera, that captures human hand skill (finger bend, fingertip force, wrist motion, and time-synced POV video) and turns it into training data for robot manipulation and embodied-AI models.

What does the GX-1 glove measure?

The GX-1 records 10 finger-bend channels (2 joints per finger), fingertip grip force on thumb, index and middle fingers, and a 6-axis wrist IMU, all at 100 Hz, hardware-synchronized and video-sync capable.

How can I get paid to capture data for BLO LAB?

Skilled people join Talika (talika.ai), wear the gear, and complete real-world tasks. Every verified hour of skilled work becomes training data and a paycheck; gear ships to your door.

All articles

Research·9 min read·August 14, 2026

The Robot Foundation Model Registry: 18 VLAs and Policies, and Which Are Open

A cleaned, sourced registry of the vision-language-action models and manipulation policies that run on robots, from 27M-parameter Octo to 55B RT-2-X. Two-thirds ship open weights; the biggest company results do not.

In 2026 a robot's application layer is increasingly a model, not a program: a vision-language-action model or policy that maps camera and language to motor actions. They are described in incompatible ways across arXiv, model cards, and GitHub, so we refined eighteen of them into one schema, parameters, type, training data, embodiment, and whether the weights are open. The clean version is at blomegalab.com/data/vla-policies.json.

The app layer is mostly open, the biggest results are not

Two-thirds of the registry ships open weights: Octo, OpenVLA, Physical Intelligence's pi0, NVIDIA's GR00T line, Hugging Face's SmolVLA, plus the ubiquitous ACT and Diffusion Policy. That is a genuinely open app layer, which is why fine-tuning a manipulation policy is now a weekend, not a research program. But the six closed models are the headline ones: RT-2 and RT-2-X, Google's Gemini Robotics, Figure's Helix, 1X's Redwood, and Dyna's DYNA-2, the model that proved the human-to-robot scaling law. The pattern from the sensor and dataset registries repeats: the open community builds the tools, the companies keep the crown jewels closed.

Of 18 robot foundation models, 12 ship open weights and 6 are closed. The app layer is open; the frontier results (DYNA-2, Helix) are not.

The registry

Eighteen models normalized to the same fields: organization, parameters, type, training data, embodiment, open weights, and year. Parameters span three orders of magnitude, from the 27-million-parameter Octo Small to the 55-billion-parameter RT-2-X. The full queryable version is the JSON below.

How it was built

Each model was pulled from its paper, model card, or company page and normalized, with confidence flags. Parameter counts are firm for the open models and marked undisclosed for the closed ones that do not publish them; where a follow-up version (pi0.5, GR00T N1.5/N1.6) reuses an architecture, it is listed separately with what changed. Nothing is invented, and closed is recorded as closed even when the model is the strongest result of the year.

Why Blomega maps this

These models are not compute-limited, they are data-limited, and every one of them, open or closed, is bottlenecked on the same input: diverse, real, force-aware demonstrations. The dataset registry showed that input is almost never force-instrumented; the sensor and force-to-task registries showed why that matters. A registry of the models closes the loop: this is the demand side, eighteen hungry consumers, for exactly the data Blomega captures.