Embodied Sim-to-Real Proof Packets

Status

This page reports a bounded Project Telos preflight, not a deployment claim. The evidence label is EMBODIED_SIM2REAL_FIXTURE_MATCH: a deterministic differential-drive fixture declares robot units, command logs, predicted and observed traces, tolerances, safety envelope, latency bound, and negative controls.

Fixture

Robot

Differential-drive planar fixture, wheel base 0.5 m, robot radius 0.08 m.

Trace check

MATCH. Mean path error 0.011756 m; terminal position error 0.021633 m.

Safety check

MATCH. Minimum obstacle clearance 0.177892 m; workspace, speed, angular-speed, and latency bounds hold.

Rejected controls

Wrong wheel base, swapped wheels, centimeters treated as meters, unsafe clearance, and latency over limit all return DRIFT.

Non-claims

No real robot safety claim, no medical or surgical claim, no foundation-model benchmark, no large-scale sim-to-real claim, and no BuildLang/buildc-native runtime receipt yet.

Why this matters

Embodied AI sits at the crossing of robotics, language, perception, control, safety, medicine, manufacturing, materials, and simulation. A fluent model answer is not enough. A serious embodied claim needs units, morphology, commands, sensors, environment state, trace comparison, safety bounds, latency, negative controls, and a verifier that can say DRIFT.

The current artifact is intentionally small. Its job is to prove the packet shape before stronger robotics claims are attempted.

Source leads

Embodied foundation models

2606.11324v1, 2505.20503v2, and 2503.20020v1 are metadata-only source leads.

Vision-language-action robotics

2307.15818v1 is a metadata-only source lead.

World models and simulators

2507.00917v3 and embodied AI survey rows pressure simulator and trace requirements.

Safety

2605.02900v2 pressures risk, attack, defense, and envelope fields.

Medical and soft robotics

Surgical and soft-robotics rows are source leads only; they do not authorize clinical or deployment claims.

Toolchain map

Gather

Captures robotics papers, benchmark cards, safety reports, datasets, and videos as receipts.

Index

Packages morphology, command logs, sensor traces, environment geometry, code, and source refs.

Forum

Routes embodied claims through robotics, safety, domain, verification, and publication lanes.

Crucible

Rejects robotics claims without units, tolerances, trace comparison, safety envelopes, latency boundaries, and negative controls.

Learn

Turns packets into exercises about units, kinematics, clearance, latency, and overclaim boundaries.

BuildLang/buildc

Target runtime for typed units, kinematics, dynamics, trace schemas, and safety-envelope checks.

Telos

Binds source, robot state, model action, environment state, verdict, and learning receipts into one proof-carrying packet.

Receipts

• Source ledger: demo/research/embodied-sim2real-source-receipts.json.
• Fixture CLI: demo/embodied-sim2real-proof-packet.mjs.
• Fixture output: embodied-sim2real-proof-packet-2026-07-02.json.
• Crucible verdict: MATCH 3 / DRIFT 0 / UNVERIFIABLE 0.
• Crucible run hash: b1f5bd65975c9a454ca9593c3b9310b9b7683ece8711ecbbcbdc789cff1f9704.
• Crucible report hash: 195c5b908d4597e38ee98dbf963d71aca564ac79fb7293f17b9a65b113d4fe2e.
• Learn verdict: VERIFIED.
• Learn witness: 258663c0dd0d647de661602ceaeb00771a1a750a478ddb562bf21c0af71c7d6a.

Next promotion target

The next public demo should be a BuildLang/buildc typed-unit replay of the same differential-drive fixture, followed by a manipulation fixture with object pose, contact state, action budget, safety envelope, negative controls, and a Learn prooflesson.

source receipts
-> typed robot morphology
-> unit-checked command log
-> predicted trace
-> observed trace
-> safety envelope
-> negative controls
-> Crucible verdict
-> Learn prooflesson

Do not infer

• Do not infer that Telos proved real-world robot safety.
• Do not infer that the fixture validates a vision-language-action or foundation model.
• Do not infer that the fixture supports surgical, medical, or clinical deployment.
• Do not infer that the fixture proves large-scale sim-to-real transfer.
• Do not infer that BuildLang/buildc already executes this robotics runtime.