Foundation Model

CFG-1

The robot foundation model, built on real data.

CFG-1 is Config’s first-generation foundation model for bimanual manipulation — trained on in-house human and robot data, and adaptable to new tasks with hours of teleop.

  • 2B

    Parameters

  • 10K+

    Training hours · in-house

  • <50ms

    Inference latency

  • 3 min

    Native memory

Approach

Human data, converted to robot data.

Most robotics foundation models scale by collecting more robot teleoperation data — expensive, slow, and operationally heavy. CFG-1 takes a different path.

  • 01 · Source

    Leveraging the scale and precision of human data.

    Most of CFG-1's data comes from humans performing tasks — far faster and cheaper to scale than teleop — mixed with high-precision robot teleop.

  • 02 · Conversion

    Translate the data, don't reshape the model.

    Our Conversion Engine maps human demonstrations into train-ready robot data at micrometer precision — the data quality of teleop, at a fraction of the cost.

  • 03 · Scale

    Built for industrial throughput.

    Our Hanoi Data Studio runs 100+ operators producing ~20,000 hours of human data a month. With teleop collection in Seoul, Config has collected 200,000+ hours across 10,000+ tasks.

  • 04 · Adaptation

    Fine-tuned to your task in hours.

    CFG-1 reaches 80%+ task success with just 1 hour of task-specific teleop data.

Architecture

Long-horizon memory.
Short-horizon precision.

CFG-1 is an autoregressive model designed for the structure of real-world manipulation: tasks that unfold over minutes, with actions that demand sub-second accuracy.

  • 3-minute native memory

    Long-horizon context window captures full task structure — pick, place, recover, repeat — without losing track of what came before.

  • Sliding-window precision

    A short window of recent frames drives high-resolution action prediction — micrometer-level accuracy where it matters, every step.

  • Edge-deployable

    2B parameters, optimized for under-50ms inference on a single RTX 5090. Runs on-prem, in the factory, without cloud round-trips.

Performance

From task description to working model.

Fine-tune, evaluate, refine — a task description becomes a deployable policy in a few iterations.

  • 80%+

    Task success rate

    Reached with just 1 hour of task-specific teleop data, fine-tuned on CFG-1.

  • 24h

    Task adaptation cycle

    Collect today, train overnight, deploy the next day.

Research

Research & publications.

We publish technical previews, evaluation notes, and field reports as CFG-1 evolves.

CFG-1, in production.

CFG-1 powers every Config Loop deployment — from feasibility test to live operation. See how it deploys on your robot.