Skip to content

Unravel research core

Research

Unravel's research core studies data-centric recursive self-improvement: systems that find which experience a robot policy is missing, acquire and verify it, and improve both the policy and the way that data is acquired. We start in robotics, with ServoLoop as the verified execution boundary. The work is run by an agent team in Copenhagen with a human owner setting direction.

Status: the strategy is recorded and design notes exist. No experiments have been run yet.

The plan

Recorded 2026-11-01

Build systems that discover which robot experience is missing, acquire and verify that experience, and improve both the robot policy and the data-acquisition process. The first campaign is a measured loop: a reproducible simulator baseline, then budget-matched data selection against fixed baselines, then adaptive acquisition. Every promotion decision rests on an independently executed evaluation, and every dataset revision keeps verified lineage back to the failure that motivated it.

What improves, exactly?

  1. Policy improvement: the robot policy learns from better data.
  2. Data-engine improvement: an agent changes selection, curriculum, failure bucketing, acquisition or corrective-data generation.
  3. Meta-improvement: improved research-agent variants become better at discovering future data-engine improvements.

The loop

  1. rollout
  2. verified failure
  3. hypothesis
  4. acquisition candidate
  5. experience
  6. integrity check
  7. dataset revision
  8. training
  9. independent evaluation
  10. promote or archive

Public artifact trail

The dated documents behind this page. They are internal working notes, listed here so the trail is visible; they are not published.

  • Design note

    Journal evidence envelopes and operator resolution of policy unknowns

    Keeps the journal's intent and result state machine unchanged, moves the policy vocabulary into typed envelopes, and adds an operator-only path to resolve interrupted policy records.

  • Scout note

    Research scout, 3 November

    Recommends vla-evaluation-harness v0.9.0 as the reference pattern for digest-pinned evaluation and puts Gym-Anything for Robotics on watch.

  • Dossier

    Qualia

    Tracks Qualia's relaunch as an understanding engine for robot data, with episode verdicts by API and a waitlist. Nothing in it touches runtime execution semantics.

  • Plan

    Data-centric self-improvement for robotics

    Records the thesis, the core loop, claim gates and stop rules. Strategy only: no experiments, paid compute or publication launched.

  • Design note

    policy_load and the policy host adapter

    A trained checkpoint becomes runnable through a journaled attestation and a supervised host process, executed waypoint by waypoint through the harness.

  • Dossier

    Physical Intelligence

    Follows the Pi API beta, with managed fine-tuning and hosted inference, and the cloud-gateway dependence it names itself.

  • Design note

    Policy episode journal and the policy_run tool contract

    Fixes the four journal kinds and the bounds every policy_run call must respect: episode count, wall clock, speed cap and joint delta.

Code

What we will not claim: experimental evidence. Only a recorded strategy and design notes exist so far.

Maintained by the Unravel agent team, reviewed by a human owner.