Voice-to-action interlock · AI Infra Summit Hackathon
The gate between hearing and hands.Nothing moves until the command passes policy.
Bollard takes every spoken command to a robot arm, transcribes it with Speechmatics, judges it against a safety policy, and only then lets the arm move. Each command leaves a tamper-evident receipt.
hear → parse → judge → act → receipt · deterministic policy · no LLM in the command path
the cockpit, live idle state · two arms, one bollard, thirteen table objects
Mechanism
Hear
Speechmatics Realtime streams your mic at 16 kHz. Partials arrive while you speak; the final transcript triggers judgment.
Judge
A deterministic grammar extracts the intent; the policy engine checks hazards, reach, grip and destination. Allow or deny, with reasons.
Act, then receipt
Allowed commands animate the arm on the dinner table. Every command appends a SHA-256 hash-chained receipt of what was heard, allowed, and moved.
What Bollard allows
- · place, pick up and slide safe objects: plates, bowls, cutlery, napkins
- · glassware with a care note: slow approach, soft grip
- · named-arm routing when you say “with the left arm”
What it refuses, with reasons
- · the hot saucepan, the lit candle, any throw of glassware
- · destinations off the table edge, occupied spots, out-of-reach corners
- · any command that arrives while the arm is mid-action
Receipts · two examples
Every command appends a receipt to a SHA-256 hash chain: the words heard, the transcription confidence, the parsed intent, the verdict with reasons, the scene fingerprint after. One click re-hashes the whole chain and says whether it holds.
#001 · ALLOW · “place a plate on the left placemat”
Policy pass: plate → the left placemat via left arm. · scene@-0.42,0.08 · sha 8e60687e…
#002 · HAZARD_HOT · “move the pan to the center”
The saucepan is hot (60°C+ on the hotplate). Manual handling only. · sha 1ecb33fe…
real receipts from a live cockpit session · hit Verify chain in the app to re-hash them