BETL (BPU-Education, Transport, Lidar), pronounced "Beetle"

A scratch-built, bio-inspired expressive lidar mobile robot on the RDK X5, built for the Robotics Dream Keeper Challenge (D-Robotics, 2026). It drives, builds 2D occupancy maps, and runs BPU-accelerated object detection concurrently — a modular research platform with personality baked in from the start.

Status: Stage 2 (Build) complete — see PROPOSAL.md and ROADMAP.md. Stage 3 (working demo) in progress toward July 15.

---

What it is

• 4-wheel skid-steer base: hoverboard hub motors, 4× Flipsky ODESC V4.2 drivers on the X5's native CAN FD bus, AS5600 wheel odometry.
• Concurrent perception: YOLOv8s object detection on the BPU + slam_toolbox 2D occupancy mapping on the CPU — the two concurrent workloads, on separate compute units.
• Forward sensing: D-Robotics stereo (BPU detection + depth) and RealSense D435i (active-IR depth + IMU), fused so their failure modes cancel. Side 8×8 ToF for obstacle avoidance.
• "Poor-man's spherical lidar": a 2D RPLIDAR A1 on a closed-loop-stepper-driven slip-ring stage, encoder-reprojected into a 3D point cloud — an optional, swappable payload.
• Modular by payload: an M6-grid cheese plate; a software-declared capability system where optional modules (lidar, cargo load cell, future sensors) enhance behavior when present but never gate the always-present baseline.
• Personality from the start: round-array LED eyes, a scrolling status panel, and a deterministic offline Speak & Spell-style voice ("Hey Betl").
• Learning on-robot (goals): a mass-conditioned energy-efficient locomotion policy and a goal-conditioned ball-pushing policy, both trained in Isaac Lab with sim-to-real calibration from real motor telemetry.

Architecture at a glance

See PROPOSAL.md for the full system-flow and ROS 2 node-graph diagrams, module design with failure modes, compute allocation, and the engineering plan (BOM, timeline, risks). Diagrams render inline on GitHub (Mermaid).

Key property: one path to the motors. Every velocity source — teleop, nav/policy, obstacle veto, cargo-drop event — funnels through a single arbitration node carrying the soft e-stop, backed by an independent hardware e-stop.

Repository layout

PROPOSAL.md      Stage 2: concept, architecture, engineering plan (+ diagrams)
ROADMAP.md       Milestones and dates through the demo
docs/            benchmark methodology, architecture notes, build journal
ros2_ws/src/     ROS 2 packages (bringup, description, drive, perception, slam,
                 lidar_pkg, safety, capability, voice, expression)
firmware/        RP2350 micros (stage controller, odometry, power, load cell, LED)
isaac/           Isaac Lab policies (locomotion, ball-push)
models/          Vendored YOLOv8s .bin + inference pipeline
tools/           Roller-dyno characterization
cad/             Frame, Y-horn, cheese plate, wheel scans

Follow the build

• YouTube (RustedFriend) — weekly recap build log (Hopefully) + the Stage 3 demo
• Instagram / Threads (RustedFriend) — visual updates
• Discord — running build log

License

TBD.