Illustrative research engagements.

Illustrative · BC startup

Bin-picking prototype for mixed-SKU totes

A Vancouver hardware startup had a vision model that worked on flat backgrounds and collapsed when liners reflected overhead LEDs. We reframed the engagement as perception plus grasping research: synthetic clutter with measured BRDF proxies, a two-finger bench arm and operator-supervised picks. After six sim cycles and two guarded hardware weeks, pick reliability improved on their SKU set but not on untested variants — documented explicitly for their investors. Human operators handled every live reset; no unattended runs.

Illustrative · deformable goods

Manipulation policy for soft pouches

Deformable objects punish naive grasp heuristics. We collected teleoperation traces, trained an imitation learning policy in sim and measured transfer on a collaborative robot arm with force limiting enabled. The policy handled nominal pouches; seam overlaps still failed at measurable rates. Client received ROS 2 nodes, evaluation logs and a sim-to-real gap memo — not a promise of line-ready automation.

Illustrative · inspection OEM

Low-light perception model for weld inspection

An equipment builder needed edge inference on a portable rig where shop lighting varies by site. We fused LiDAR depth cues with a compact vision backbone, ran digital-twin glare scenarios and validated on hardware with CSA-aware guarding on the test stand. Model met agreed precision on held-out seams; recall dropped when operators wore reflective vests — a finding that shaped their deployment checklist rather than being buried.

Evaluation snapshot

Sim hours: 840 · Hardware sessions: 11 · Operator-supervised throughout · No guaranteed field accuracy

Illustrative · logistics pilot

AMR perception stack for narrow-aisle pallets

Autonomous mobile robot perception in tight aisles combines SLAM drift, occlusions and pedestrian traffic. We prototyped a sensor-fusion stack with teleoperation fallback, logging near-miss events for review. Pilot ended with clear go/no-go criteria for a wider fleet trial — not a purchase order for autonomy.