Collision avoidance that obeys the rules of the road: velocity obstacles against rule-based and sampling MPC for a small boat

What it found

Milestone 1 of track 23's marine-colregs step, filed under the marine-hull programme. Kinematic own boat (4 m/s, 6 deg/s turn limit) against a target on constant course (it never gives way). 115 scripted head-on, crossing (both sides, speed ratios 0.5 to 1.5), overtaking and overtaken encounters with lateral offsets of -100 to 100 m. Roles are classified from Rules 13 to 15; planners: rule-based 35 degree starboard turn, velocity obstacle limited to starboard for give-way, and sampling MPC. Scores: closest point of approach (safe = at least 100 m), rule compliance (starboard turn, pass astern in a crossing, stand-on holds course until the action range) and detour. Check: straight-course CPA matches the analytic offset. Sweeps: target-velocity error 0, 0.3, 0.6 m/s (3 seeds) and stand-on action range 200 to 600 m. Preliminary: no vessel dynamics beyond a turn-rate limit, the safe distance and speeds are chosen not sourced, the target never manoeuvres, no AIS replay, and the MPC cost weights are untuned (no CMA-ES arm). Simulation only; research, not a certified system.

Paper, code and results

• README and method
• Experiments and results
• Code
• Tests

More in Open-source purpose-built robots

• Which body carries a survey payload furthest per kWh? Hull choice for small autonomous boatsResearch programme (proposal stage): energy per survey kilometre of hull types for 100 to 500 kg autonomous boats; with a Michell wave term a catamaran out-ranges a monohull in 3.3% of 6,561 cases, none at equal length; with Slocum coefficients from the gliderflight package a glider never out-ranges a same-mass propeller AUV at equal hotel load (0 of 2,430 cases); it breaks even when the AUV's hotel load is about 5 times its own; a REMUS-100-class AUV's non-propulsion load implied by vendor endurance is 104 to 175 W, far above that, so the glider out-ranges it in every case on energy per metre (at a tenth of the speed); a Savitsky planing hull reaches at most 0.14 of the monohull range (0 of 48 cases); a SWATH (two submerged torpedoes on struts) out-ranges the monohull in none of 4,374 swept pairs; charging panel mass to the battery (2 to 10 kg/m2, swept) cuts the median solar range multiplier from 2.01 to 1.64 at worst, and the full deck stays the best array wherever range is finite; against ten years of daily NASA POWER insolation a threshold speed policy (slow or hold when the pack is under 30%) never raises km/day over a fixed 1 m/s but cuts pack-empty days, e.g. 25.4% to 1.5% for a 4 m boat at 25 N
• Holding a survey line in current and waves: line-of-sight against integral line-of-sight for a small boatPlain line-of-sight guidance leaves a steady cross-track offset of delta*tan(asin(c/U)) in a cross current c (simulation matches it to 1%: 3.77 m at 0.3 m/s, 12 m look-ahead); integral LOS removes it, with lower rms cross-track error in 136 of 144 cross-current cases and the 8 exceptions all at a 6 m look-ahead with integral gain 2, where it oscillates (rms about 2.3 m even in still water)
• If a late field loses yield instead of failing a hard deadline, does a scout-and-treater fleet still lose to inline weeding robots?Replaces the hard deadline of osrobot-deadline with a 2% yield loss per leaf stage of delay (Knezevic et al. 2003): the best pipeline costs 1.06 to 1.20 times the best inline fleet at 0.72 ha and 1.49 to 1.51 at 0.18 ha, and at 1.62 ha inline is cheaper when the window closes by 8 h and the pipeline from 12 h
• What does a 2-joint feeding arm cost against a 6-joint one once the whole bill of materials is priced?What does a 2-joint feeding arm cost against a 6-joint one once the whole bill of materials is priced? About 1.8 times, not 3: $66 against $122 at the SO-ARM101 table's $13.89 servo, rising to 2.4 times at a $49 retail servo.
• Does treating weeds as you drive still beat scouting first when the robot uses published laser-weeder numbers?Re-running the osrobot-field weeding model with sourced laser-weeder constants (2 m swath, 0.73 s per weed, 99.2% kill) halves the time but keeps the ranking: inline treatment takes 71.7 min against 79.3 for scout-then-treat at 0.3 weeds/m² and 1 km/h, and scout never wins across densities, speeds and detector hit rates
• Can a stdlib static linter catch the lab-protocol faults the Opentrons simulator accepts?Can a stdlib static linter catch the lab-protocol faults the Opentrons simulator accepts? It flags 362 of the 522 seeded faults to the simulator's 270, adding reused tips, tips left on, sub-minimum volumes and carry-over; stylised, partly circular.

All 44 Open-source purpose-built robots projects →

Other research topics

Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.