Case 06 · Industrial autonomy OEM
Edge perception mesh
On-robot multimodal perception and planning for hazardous environments — 62% faster missions, fully offline when the uplink dies, three years fielded without a safety recall.
Context
An industrial OEM builds robots that enter places humans should not: chemical plants, collapsed structures, high-heat inspection corridors. Connectivity is intermittent by definition. A cloud-dependent stack is a non-starter. Prior perception stacks froze or hallucinated clear paths when sensors disagreed — both failure modes are unacceptable.
The problem, precisely
Fuse LiDAR, thermal, and RF into a coherent world model on constrained compute. Plan under uncertainty with safety envelopes that degrade gracefully. Prove — to insurers and customers — that autonomy will refuse unsafe action rather than invent confidence.
Graceful degradation beats heroic autonomy. When sensors conflict,
slow down or stop. Never “average” your way through a hazard.
What we built
Quantized vision transformers for onboard perception, a model- predictive controller for motion, and formalized geofence monitors that could halt actuators independent of the planner. The mesh synchronized multi-robot observations when links existed, and continued solo when they did not.
Architecture highlights
- On-device multimodal fusion with conflict detection.
- MPC planner with explicit uncertainty budgets.
- Independent safety monitor with formal geofence checks.
- Offline-first operation; cloud as optimization, not dependency.
- Field telemetry for post-mission learning without live uplink.
The hard parts
Thermal and LiDAR disagreement in smoke and dust created pathological cases. We trained conflict classifiers and forced the planner into restricted modes rather than blending incompatible geometries.
Certification narratives required evidence packs, not demos. We instrumented every halt reason and shipped insurer-readable reports as a product feature — which later became a sales asset for the OEM’s own marketing team.
Rollout
Hardware-in-the-loop, controlled yards, then customer sites with escalating autonomy permissions. Three years in field without a safety recall related to the perception/planning stack.
Outcomes
Mission time down 62% on the benchmark routes. Full offline operability as a contractual guarantee. The safety evidence pack shortened enterprise sales cycles — proof that deep engineering can be a go-to-market advantage, not only a cost center.
What we’d repeat
Design the refuse path first. Make evidence a deliverable. And let marketing sell what safety engineering already proved — never the reverse.