Loading…
missionSTACK / swarmREFLEX3 + ROS
Compact decision models inside explicit contracts. Check the input. Bound the action. Keep the trail. Simulation research.
Generation 3 / Fleet & work-cell model cards
Fleet / Simulation
Keep routine decisions local.
Separate tests: coordination + rule reflexes reduced calls; model-v4 timing. Current fleet artifact: 2.1. Concept image; simulation evidence.
Fleet model & evidence →
Arms & grippers / Kinematic simulation
Bound the response at contact.
missionROS task flows and recorded replay. Hardware performance and an advantage over tuned rules remain unproven. Concept image; fleet figures do not apply.
ROS contracts & evidence →
swarmREFLEX3 / Fleet artifact 2.1 / Trained 4 October 2026
Fleet artifact swarmREFLEX-2.1 uses per-contract XGBoost models with 129 observable features. Calibrated admission thresholds bound learned actions; a named rule handles abstention.
bcebc14a…2792885bb590…0376Published simulation evidence / Separate experiments
Coordination plus rules reduced calls. An earlier model path measured latency. Separate experiments; neither measures artifact 2.1 or ROS hardware.
“Up to 90%” is a conservative presentation of the 98.3% simulation result. This is a system-level reduction, not the learned reflex model acting alone. Fleet call reduction and decision-path latency use different experiments and denominators.
Historical H-RX9 / 22 September 2026 / Simulation
swarmREFLEX-1, Jev and Laya: 300 held-out decisions, a 50 ms deadline and fallback for rejected actions. This does not evaluate version 2.1.
| Model & deployment path | Admitted | Harmful / admitted | p50 latency |
|---|---|---|---|
| swarmREFLEX-1 (v4 per-family trees, CPU) | 18.0% | 0/54 | 24.8 ms |
| Laya (open weights, root, GB10) | 0.0% | 0/0 | 30.8 ms |
| Laya typed-decisions (open weights, GB10) | 0.0% | 0/0 | 42.3 ms |
| Jev (TypeSafe System One, jev-latest, API) | 0.0% | 0/0 | 676.9 ms |
swarmREFLEX-1 admitted 54/300 decisions with zero observed harmful admissions; the upper 95% harm bound was 5.4%. Zero admissions provide no evidence of safe useful action. CPU, GB10 and remote API paths differ: this is a task-and-gate comparison, not a hardware-controlled speed test or a universal ranking.
Read the limits with the result
Read the labels, fallback and test conditions with the result. A calibrated gate is not a safety function.
Blocked-zone and charger labels agreed with forced simulation reruns at about chance level in the earlier fidelity probe. Harm bounds are relative to the declared oracle and fallback, not realised physical outcomes.
The 0.5% upper-95 calibration target and 80% cal-B admission rate describe threshold selection. They do not establish deployment safety or held-out operational performance.
swarmREFLEX3-ROS has six synthetic work-cell rule contracts. Kinematic hold-contract comparisons are published; no trained-model advantage over tuned rules or real-arm performance is established. swarmREFLEX-2.1 is not a safety function.
swarmREFLEX3-ROS / Work-cell model card / Simulation
Six rule contracts bound human reach, grip force, unexpected contact, slip, tool-process signals and placement clearance. missionROS sequences the task; the reflex controller supplies a bounded override.
Continue, slow, hold, retract, release or abort the skill, within contract-specific limits. The plain controller owns the control-tick path; LangGraph hosts orchestration, simulation and recorded replay.
MR-4e compared a trained hold contract with baseline and shortened-clear-time rules in kinematic simulation. The tuned rule was faster but recorded three true in-reach motions versus two for the model. Comparative safety remains unresolved.
Synthetic signals and simulated workers do not demonstrate sensing coverage, real-arm response time or gripper performance. Hardware integration and commissioning need their own evidence. The fleet call-reduction and latency figures do not apply to ROS.
missionSTACK / Transparent, logged & versioned
Coordinate with swarmCORTEX. Sequence with missionROS. Bound responses with swarmREFLEX3 and swarmREFLEX3-ROS. Keep context and evidence with the run.
The service validates input freshness and typed actions. Stale inputs abstain; explicit fallbacks retain control.
Decision, contract and outcome interfaces expose the action source. A hash-chained journal supports verification of the recorded decision sequence.
Replay failures, revise a contract or model and evaluate the candidate. Keep the model hash and configuration with each result.