Physical intelligence, learned from humans.
Useful questions.
Learn from data.
Test what transfers.
We begin with physical-interaction models and evaluation code, using small experiments and free, open tools.
Our near-term plan uses public data whose terms permit the work to train a focused model, then test it against simple alternatives on data kept apart from training. Potential users include large AI and robotics labs, robotics startups, and industrial robot teams in manufacturing, fulfillment and packaging. Their feedback can help us choose questions that matter.
Human recordings paired with calibrated physical measurements remain a planned way to address gaps in public data. The capture thesis is: Cr4dle turns consented recordings of how people look, grip and act during physical work into the data and models that let robots learn from humans and work beside them. The current capture demo is synthetic; real-sensor accuracy, robot-task gains and paid demand remain unproven. The end-state is a test, not a claim: the human-parity trials are a protocol — a published task list, blinded judges, published either way. A benchmark result is about behaviour, never about inner states, and it is scored in simulation, labelled so, until a real body runs it. In the repository the trials are a protocol block and a dated bet with a kill criterion.
Measured by design.
Published either way.
- 01
Sequence by evidence
force, then gaze, then state, then models; nothing moves up the order without a real-robot number behind it.
- 02
Every dataset card prints sync error, participant count, held-out-user error, discard counts and what was redacted
- 03
Sell the channel, never the hour
- 04
Buy sensors until a paying customer proves what the bought sensor cannot do
build only that.
- 05
Raw video and poses ship redacted
biosignals and gaze ship as features; consent is checked at capture and again at export, with revocation, on every episode.
- 06
Cross-user generalization is the hard problem
the LOSO harness is the acceptance test.
- 07
A dated gate and a kill criterion are written before every experiment
the result is published whichever way it falls; we replicate before claiming and cite the paper before the pitch.
- 08
The end-state is a test, not a claim
an agent's likeness to a person is a benchmark result about behaviour under a published protocol with blinded judges, never a statement about inner states, and it is scored in simulation, labelled so, until a real body runs it.
Three pillars, one mission.
Stated as the company vision states them. Each pillar is held in the repository as a dated bet with a kill criterion; the ceiling under each is where the evidence stands, not where the story ends. Small model experiments begin the research; later products and robot capabilities need their own evidence.
- World model
Research into physically grounded and human-conditioned world models, used first as evaluators and data engines; persistent-state and physical-correctness evaluation. Proposed Cr4dle models would condition on consented human-side force whose labels carry a held-out-user error; physical task benefit remains unproven.
research/,benchmarks/, laterplatform/models/Held as
wm-human-conditioned-latentinhorizons/bets.yaml: a dated bet with a kill criterion.- Physical / human model
The Cr4dle Human Model (what a person reaches for, when, how hard; behaviour near machines) and the contact model of the hand; cross-user calibration as the hard problem.
platform/decoders/human-model,sdk/pythonHeld as
human-model-paired-corpusinhorizons/bets.yaml: a dated bet with a kill criterion.- Robots
The embodied agent (planner, policy, dialogue, memory) behind the SDK; the human-parity trials as a public benchmark; bodies bought (none is bought today: the SO-101 rig is a candidate purchase in the equipment registry; humanoid platforms by gate) plus an own-platform track that is a simulation twin until its build ADR.
agent and policies:
research/now,platform/models/later (sdk/*is the interface and mock, ADR-0105);benchmarks/;hardware/platforms/Held as
embodied-agent-human-parityinhorizons/bets.yaml: a dated bet with a kill criterion.- Rung one: humans as the instrument
Force, gaze, hand/head pose, demonstrator state and consent on one clock; the acceptance test that certifies the episodes.
platform/*,devices/l0-ember,devices/l0-lens
Research now.
Capture when useful.
What is planned for a later phase: collections of consented, labelled episodes on your task (quoted per usable demonstration), the kit licence so your own demonstrators produce them, and dataset licences — each behind its dated gate. Cr4dle is a non-commercial research programme now, and nothing is offered for sale today. Ember is the one band Cr4dle plans to build. Its build gate is Q2 2027, and the build needs three paid collections on bought bands, so it waits on the decision that replaces ADR-0136. Lens is a bought pair of glasses used as shipped. Lumen and Crown are EEG research tiers for bounded sessions, and physiology is a recorded research column, never a product.
Research is the current priority. The plan for a later phase is a data offer to robot-learning teams: training data, tasks and rubrics (ADR-0136). We still need to establish usefulness and demand. The planned products keep their dated gates, and each paid gate waits on a new decision that replaces ADR-0136: none is selling today. A delivered collection still precedes capture-kit licensing. The Human Model product retains its separate licence-revenue gate; small research experiments do not establish that product is ready. Later wearable and agent stages need measured retention or physical trials as appropriate, with no promised year for the later agent stages.
- 1aInstrumented human demonstrationGated · the path exists; nothing is sold before a decision that replaces ADR-01364 dated gates, first
- 1bOperator state for robot-data operationsProposed · a row with a dated gate; nothing built for sale1 dated gate, first
- 1cCapture-kit licenceProposed · a row with a dated gate; nothing built for sale1 dated gate, first
- 2Cr4dle Human ModelProposed · a row with a dated gate; nothing built for sale1 dated gate, first
- 3Everyday wearableProposed · a row with a dated gate; nothing built for sale1 dated gate, first
Bought bodies.
Bets with kill dates.
In the words of the company vision: what is bought, what is bet on with a kill date, and what is not claimed. The vision’s bullet on workplace monitoring is not repeated here; the channel descriptions carry it: labels are fatigue and safety-relevant physical state, never a mental state. The one line is Physical intelligence, learned from humans.
- Not a robot-body or humanoid maker first.Bodies are bought until a gate says otherwise (none is bought today: the SO-101 rig is a candidate purchase in the equipment registry; a humanoid when a bet stage names one); the own-platform track in
hardware/platforms/is a simulation twin until its build ADR. We plan to sell into every robot program in a later phase (ADR-0136). - Not a generic video world-model lab.No pretraining run of a generic interactive video model; human- and force-conditioned latent models, evaluators and adapters on open weights, under research bets with kill dates; a pretraining run only behind licence revenue or a funded compute line.
- Not an hours vendor.We compete on contact truth from bare hands, with calibration and models, not on the count of channels or on hours.
- Not a neuro or physiology company.EEG, fNIRS, EDA and HRV are research tiers; the clinical research families stay as reference code and are not part of the company story.
- Not a consumer hardware company yet.Measured retention before any scale spend.
- Not a raw-video pipeline.Video and poses ship redacted; biosignals and gaze leave the device only as features; consent is checked at capture and again at export.
- Not claiming novelty.ForceBand, Feel the Force, Ishida, Gaze2Act and earlier gaze-conditioned imitation are prior results we replicate on a customer's task and report either way.