OmniLinkStart free ↗Let the robots coordinate.
Picker requests an empty cart. Tug parks and reports its measured position. Picker loads the parts, then asks for delivery to shipping.
OmniLink × OmniSim / Demonstrations & evidence
Long tasks.
Measured outcomes.
One order across three robots. A warehouse shift with rules to remember. A robot you can interrupt. See the work, inspect the outcomes, and take the tests with you.
Simulation demonstrations and historical studies with explicit scope. OmniKey and active model provider required for connected AI; usage separate. Physical robot operation is unverified.
3robots coordinating one order
79completed crew passes / 100 planned
30 minwarehouse commitment study
198matched robot-control trials
One order · Picker, Tug & Packer
Bring the cart.
Pack the order.
Ask for one red, one blue and one green part in the shipping crate. Two OmniArm 6 robots and an OmniTug coordinate the handoffs, move the parts and report the result.
Close the loop.
Packer unloads the three parts into the crate. The order counts only when the final scan places every part in the crate, Sam receives a completion report, and no delivery error is recorded.
Test another fresh session.
79/79 completed runs passed, with 610 crew messages. One proposed distance was refused because no one had specified it; the crew recovered using the named destination.
Historical campaign, 21 October 2026, Sonnet 5.5 with standing robot briefings, eight sessions in parallel. This developer-known task followed earlier pilots and fixes; all are documented. Vacuum grasps are simulated welds and the trolley follows a velocity servo, rather than a modelled drawbar joint. No per-run machine fingerprint was retained, so no throughput claim is made. The results do not establish a general safety or hallucination rate.
Crew logs, final scans & offline verifier ↗Fresh offline regrade ↗Original campaign summary ↗Earlier failures & fixes ↗
# Extract crew-evidence.zip. No network or credentials are needed to audit it.
python -I -S verify_crew.pyOmniArm 6 / Text and typed controls
An arm that waits.
An outcome you can check.
Ask for an ordered route through inspect, transfer and home. Hold it, quote an instruction, ask a question, request an out-of-cell target, then cancel a running plan. Useful work and unwanted motion are scored separately.
220live requests / 20 fresh sessions
40/40ordered route completions
0/140unwanted-motion violations / designated checks
0/98unsupported / unique published step receipts
19/20 episodes passed; 1 failed and 0 invalid/error episodes. 19/20 Cancel/Stop requests reached a running plan; 19 passed the halt criteria. One plan completed before its start reply returned, so the Cancel fixture remains a failed primary case. Separately, 3/3 cancellations during a pending HTTP request passed, without replacing the primary results.
Keep the operator in control.
Explicit Hold, Stop and Resume have deterministic handling. A hold persists through unrelated requests and quotations. Resume clears the hold without restarting cancelled work. Independent typed Stop remains available without AI.
Measure the outcome.
A step receipt requires measured joint arrival and rest. Outcome text comes from verified state rather than unrestricted model prose. 0 unsupported receipts were found among 98 unique published step receipts in this suite; that is not a general model hallucination rate.
Make the boundaries testable.
The model proposed 30 tool calls that were refused. Separate controlled checks cover 100 deliberately bad proposals, five useful motion controls and three independent typed Stops. Injected checks are reported separately from live requests.
Machine 9722d23d12a3, Newton / MuJoCo CPU, 16 ms and eight configured substeps. Relay origin and retained model traces establish connected model use; command parsing was disabled for qualification. This developer-authored regression suite is not an independent blind holdout. HTTP text and typed controls are qualified; native-window language and voice are not. No obstacle perception, collision-avoidance, force, payload or hardware-safety qualification is claimed.
Development misses and invalid attempts are retained, including the explicit Unlock omission found before the final repair. Every campaign, correction and model condition is listed in the report. The observed rates describe this finite contract and these requests; they are not a promise about arbitrary language or universal safety.
Report, criteria & development history ↗Results JSON ↗Every case CSV ↗Raw evidence + offline verifier ↗Rerunnable source add-on ↗Source SHA-256 ↗
# Working public OmniSim checkout + source add-on extracted at its root.
# Connect your own OmniKey and active model provider; set OMNI_KEY privately.
python -m omnisim doctor
python -m pip install psutil
# Reuse your own dedicated profile for serial local reruns.
$env:OMNILINK_AGENT_TAG = "my-arm-assessment"
python arm_assurance.py run --episode 0 --out .tmp/arm-live-01
python arm_assurance.py verify .tmp/arm-live-01
# Full qualification: follow the package guide for dedicated parallel profiles.
python arm_assurance.py suite --parallel 1 --out .tmp/arm-suite-01Use a fresh output directory. OmniKey and your active model provider are required; AI usage is separate. No credentials are bundled. Windows has the only prebuilt OmniSim package; other platforms build from source. Fifty-three arm unit checks and 55 relay checks exercise failure handling separately from the physical campaign.
Published protocols · retained failures
Beyond the
selected take.
A demonstration shows one outcome. A study records the workload, conditions, all attempts and scoring rules. Use the published evidence to assess the fit for your own robot work.
Warehouse operations / 30 minutes
Keep commitments as the shift changes.
Standing restrictions, delayed tasks, conflicting requests, disturbances and Stop. The original OmniLink configuration kept 89.9% of scored commitments across three repeats with Gemini 3.5 Flash.
33 scored checks per repeat, seven original configurations, machine 9722d23d12a3. The study includes keyword-graded replies and real-time slowdowns. Later Codex and Claude Code runs used different models, dates or builds; they are retrospective system comparisons. Read those conditions before comparing scores.
Robotics control / matched model
Compare the same tasks and settings.
198 trials: 33 tasks across Husky, TurtleBot3 and OmniArm 6 for six integrations, all with Opus 5.5 at low effort. Five integrations, including OmniLink, passed all 33 tasks. Plain tools passed 32.
Previously seen, developer-authored development suite; one repeat per task. Assisted manipulation remains assisted. These are specific integrations, rather than complete hosted products or a market-wide ranking. No evidence of universal capability leadership is implied.
Inspect every integration’s task count
| Integration | Passed / attempted | Unwanted-motion flags |
|---|---|---|
| OmniLink | 33/33 | 0 |
| Plain tools | 32/33 | 0 |
| LangGraph | 33/33 | 0 |
| Lobster | 33/33 | 0 |
| LangGraph + assistance | 33/33 | 0 |
| Lobster + assistance | 33/33 | 0 |
Deep Robotics X30 / Plant inspection
Ask for the task.
Resolve the ambiguity.
Send the quadruped to the feed pump. “Go to the tank” needs clarification because the world has two tanks. Change the plan, interrupt a walk, then return to the dock.
Selected native simulation take, mixed command and model paths. Known walkway graph and simulation position; no obstacle sensing and no gauge readings. Flat-ground scripted trot on contact physics. Waiting trimmed; some walking plays at 2×. No hardware performance claim.
A useful agent comes with its acceptance tests
Your robot.
Your operating rules.
Tell us what the robot must achieve, what it must refuse, when an operator must be able to stop it, and what evidence counts as completion. We build the workflow and deliver the tests with it.
The Husky mission pilot already includes ordered station visits, cancellation, independent Stop, measured arrivals and a rerunnable handover.
Robot Agent Build
Agree the work.
Keep the evidence.
- Robot integration + agreed operations
- Defined action and interruption checks
- Measured report + source + rerunnable tests
- Acceptance criteria before delivery
Scoped quote. OmniKey and your active model provider required; AI usage separate. Demonstrations here are simulation only. Hardware integration and its safeguards are scoped separately.