Roadmaps / Robotics / The milestone project
Read first
The milestone project
Scenario. A warehouse has a 30 m by 20 m floor with shelf aisles 1.5 m wide. A delivery robot (circular body, 0.4 m across, top speed 0.5 m/s, two driven wheels) carries parcels between 20 named points. People walk through the aisles. The target is to reach a goal within 0.15 m and 0.2 rad in at least 95 of 100 runs (20 goals, 5 random seeds), never touch an obstacle, and stop for a person who steps in at 1 m. A lost robot, a late sensor or a crashed node must end in a stop and a report, never in a robot that keeps driving.
| Phase | Checkpoint | Done when |
|---|---|---|
| 1 | Simulated robot | The robot is described in URDF, runs in Gazebo and is driven by ros2_control with a differential-drive controller. The transform tree matches REP 105 (map, odom, base_link) and a script checks it. A 10 m square driven by teleoperation ends within 0.5 m of the start under the simulator's wheel noise, and a report gives the command-to-wheel delay (median and p99 over 1,000 commands). A seeded wrong frame name is caught by the tree check. All runs are saved as bag files. |
| 2 | Sensing and mapping | The robot has a 2D lidar and an IMU. The state estimator fuses wheel odometry and the IMU, and over three repeats of the 10 m square its end-point error is at least 30% smaller than wheel odometry alone (a table gives both). slam_toolbox builds a map of the world, the map is saved, and a map-versus-truth overlap score is reported. A seeded featureless corridor makes mapping drift, and the report shows it and what you changed. |
| 3 | Autonomous navigation | Nav2 drives the robot to 20 goals with 5 random seeds each (100 runs). At least 95 reach the goal tolerance with no contact. Every failure is classed (planner, controller, localisation, stuck, timeout) in a table. A moving obstacle that appears 1 m ahead makes the robot stop or go around, and the stopping distance is measured. (Fresher: 10 goals, 3 seeds, and report the number you got.) |
| 4 | Safe and reliable | A command watchdog zeroes velocity within 0.3 s of the last command, and a collision monitor stops the robot for an obstacle in a 0.6 m zone. Over an 8-hour simulated soak (fresher: 2 hours) with injected faults (lidar dropout for 5 s, a killed node, 200 ms of added delay) the robot never touches an obstacle and recovers or stops in every case. A launch test and a bag-replay test run in CI. The report separates what the simulation shows from what only hardware can show. |
| W | Work version (Staff and Principal) | On a real robot or navigation stack: measure a baseline (mission success rate, interventions per 100 operating hours, stopping distance), make one change (a controller retune, a watchdog, a localisation fix) and measure again. Pass: a target agreed in writing beforehand is met, or interventions fall at least 30% on the same missions. The ADR is signed by a named stakeholder, scored by someone other than you, and a recheck is booked three months later. Level 4 also needs evidence that another team used your checklist and recorded a finding with it. Timebox: six weeks alongside Phases 3 and 4. |
Terms. A frame is a coordinate system attached to something (the map, the start point, the robot body). A transform says where one frame is relative to another. Odometry is the robot's own estimate of how far it has moved, usually from wheel turns; it drifts. Localisation finds the robot's position on a known map. SLAM (simultaneous localisation and mapping) builds the map while finding the position in it. A costmap is a grid of how costly each cell is to drive through. A recovery is an action the robot takes when it is stuck, such as backing up.
Evidence to hand in. A repository whose README opens with the problem and the numbers above; one ADR (architecture decision record) per decision, each naming a rejected option; bag files and a test report for every checkpoint; a one-page design doc; a diagram of nodes and topics. Skeletons: Checkpoint 1 is the tree check output and a delay table. Checkpoint 2 is the drift table (run, odometry error, fused error) and the saved map. Checkpoint 3 is a 100-row run table (goal, seed, result, failure class) with the bag for each failure. Checkpoint 4 is a fault table (fault, expected, observed, stop distance).
| Level | What you build and hand in |
|---|---|
| Fresher | Core scope: an empty world, 10 goals, a 2-hour soak, the four checkpoints, ADRs 1 to 4. |
| Senior | Adds a cluttered world, moving obstacles, the fault suite in CI and a C++ watchdog comparison, plus ADRs 5 and 6. |
| Staff | Adds written decisions with numbers: a safety case (hazards, controls, residual risk), a failure taxonomy with rates, and what the operations team must monitor. |
| Principal | Adds a platform and safety strategy for a fleet, a review checklist other teams reuse, and a measured outcome from the work version. |
Do this at work. A solo lab is not promotion evidence. Take a real robot or autonomous system that needs interventions, and run Checkpoint W on it. Start with a baseline: success rate, interventions per 100 operating hours, the three most common failure causes and the stopping distance. Make one change. Measure again. A promotion committee accepts the real decision and who made it, the named stakeholder who relied on it (for example the operations lead), before and after numbers, and a result checked again three months later.
Solo version. With no real robot, use Nav2's own simulation bringup and its documentation as the baseline stack, and measure your changes against it. Ask a peer or mentor to review your run tables and failure classes. Replay recorded bag files instead of live robot data.
Laptop scale. The whole core path runs in simulation on a laptop. The 8-hour soak runs overnight, and a faster-than-real-time simulation shortens it if your machine allows. Nothing needs cloud spend.
Honest sizing. A robot that drives to one goal takes an afternoon. A robot that reaches 95 of 100 goals and stops safely on every injected fault takes weeks, because the last failures are rare and need recordings to find. Phase 3 and Phase 4 carry the most load for a fresher, so use the extra fresher week there.
Worked reference: what a "3 Proven" hand-in for Checkpoint 4 looks like
The numbers are illustrative. Run your own faults and hand in your own output.
| Injected fault (robot at 0.5 m/s) | No safety layer | Watchdog and collision monitor |
|---|---|---|
| Controller node killed | Robot keeps its last speed | Stops within 0.3 s, travels 0.15 m |
| Lidar silent for 5 s | Drives blind, contact at 2.1 s | Stops after the lidar timeout, no contact |
| Person steps in at 1 m | Contact | Stops, 0.35 m travelled, 0.65 m gap |
| 200 ms extra delay on commands | Overshoots goal by 0.3 m | Within tolerance, slower approach |
ADR-6 excerpt: Stop zone of 0.6 m, command timeout of 0.3 s. Stopping distance is reaction distance plus braking distance. At 0.5 m/s with 200 ms of latency the robot travels 0.5 x 0.2 = 0.10 m before braking starts. With a deceleration of 0.5 m/s squared it then needs 0.5 x 0.5 / (2 x 0.5) = 0.25 m. Total 0.35 m, so a 0.6 m zone leaves 0.25 m of margin. A longer timeout gives more travel after a crash (0.5 x 0.3 = 0.15 m), a shorter one risks stops from normal jitter, so 0.3 s sits above the 99th percentile command gap you measured in Phase 1. Rejected: stopping only in the controller, because a crashed controller cannot stop itself. Rejected: a larger zone, because the 1.5 m aisle then blocks the robot from passing a person at all. Consequences: the real robot needs its own hardware emergency stop, since none of this survives a hung computer. Revisit if top speed or measured latency changes.