OpenEmbody

POLICY · CODE · AGENT · MEMORY

OpenEmbody

A workspace for embodied intelligence.

From demonstrations to actions. From outcomes to the next experiment.

Open console Learn more
Platform concept illustration
Invited research previewBuilt around observable outcomes

THE PLATFORM

One workspace.
Every step of the robot learning loop.

OpenEmbody brings together policy training, checked code skills, agents, and source-linked memory. Connect your data to experiments, inspect what happened, and use the evidence to decide what comes next.

POLICY BACKEND
VPP2
ENVIRONMENT
RoboDojo simulation
WORKFLOW
Train · Evaluate · Review
ACCESS
Invitation only
Documentation ↗Enter workspace ↗

DEVELOPMENT WORKFLOW

From demonstration to iteration.

A shared record of data, model versions, execution, and evaluation.

01

Prepare data

Define the robot and action contract. Validate demonstrations and retain their source and outcome labels.

DATA & PROVENANCE
02

Train a policy

Run a budgeted training job. Inspect progress, checkpoints, cancellation, and cleanup evidence.

MODELS & CHECKPOINTS
03

Evaluate actions

Compare a candidate with its baseline in simulation. Review original camera recordings and measured outcomes.

RUNS & EVALUATION
04

Review and improve

Use code skills, agent feedback, and memory to guide another experiment. Keep release decisions explicit.

AGENTS & MEMORY

LOCAL EVALUATION

Outcomes you can inspect.

Diagnostic results from October 20, 2026 · RoboDojo stack_bowls

POLICYEPISODESSUCCESSCONTROL STEPS / EPISODERELEASE STATUS
VPP2 baselinejoint2b_s10000020 / 2800Reference
Output-head candidate2 diagnostic training steps20 / 2800Not approved

Two seed pairs test the pipeline; they do not establish statistical improvement. Original recordings are available to invited project members.

Inspect evaluation ↗

CURRENT CAPABILITIES

Build on what is already working.

POLICY

Real model inference

VPP2 action prediction and isolated candidate loading have been verified in simulation.

Verified locally
TRAINING

Traceable post-training

Real output-head updates, checkpoint reload, cancellation, and deadline handling.

Fixed diagnostic recipe
AGENT + HARNESS

Evidence-linked execution

Checked tool execution, source-linked memory, and external agent assistance.

Partial integration

Independent benchmarks, full assisted self-improvement, additional model backends, and physical robot execution remain in development.

DEVELOPER ACCESS

Start with an invited workspace.

Resource API

Project-scoped read keys are available. Public policy inference is not yet enabled.

GET /api/v1/models?project_id=…
Authorization: Bearer <YOUR_READ_KEY>

Client downloads

Windows and Debian clients are planned. Installable releases will appear after build and installation validation.

No installers released yet

UPDATES

The invited web preview is available with project access, training records, model candidates, read keys, and protected evaluation videos.

Read the guide ↗