The lab

Paper 2 · Asymmetric-Information Predictive-Intent Release for Cooperative Driving

Who needs to know?

Draw a line down the street and the blue car drives it through recorded traffic. Twice a second it works out which neighbours actually need to know where it is going, and those give way. Broadcasting to everyone in range is just as safe and costs several times the bytes.

Lankershim Blvd · 0 vehicles
0.0 s
Typed release
Draw a line and a car drives it — roads optional
  • No line. Nothing that car could still do meets your plan, so it is never told.
  • Teal, 6 bytes. Everything it could do meets your plan. It gives way on those six bytes; your trajectory would add nothing.
  • Amber, 30 bytes. Some of its options conflict and some don’t, so it needs your actual trajectory to decide.
  • Faded ring. It read your plan, found no conflict and carried on. Most links look like this.

Try this

Drive on Broadcast everything and watch the lines fan out to every car in range. Drive the same line on Typed release: the same cars brake at the same moments, but most of the lines are gone. Then try Say nothing.

Across all 279 starts here, broadcasting costs about 20,056 bytes a run against 3,484 — a factor of 5.8. Collisions run 25% silent, 13% broadcasting and 14% typed.

Most cars are never told

Little seems to happen, and that is the result rather than a flaw. The rule stays silent toward 73% of the neighbours it considers, because every route still open to them is clear of you, and only about 21% of the vehicles that are told ever give way.

So the radio changes how 18% of runs here end and saves 12% of them from a collision. Not all of them: a car told 20 m from a junction at 11 m/s cannot always stop.

Why the silence is safe

Your car goes quiet toward a neighbour only when every route still open to it is clear of your plan, with the conflict test inflated by the 4 m map-match and sampling tolerance — the paper’s sound configuration, under which a missed conflict is ruled out by construction.

Across the full recorded corpus it missed 0 of 84,208 conflicts at 3.83× fewer bytes than broadcasting. The selector deciding on this page is that configuration, and it matched the Python reference on all 28,792 recorded decisions.

Why it matters at fleet scale

One car saving bytes is a curiosity. Every vehicle that broadcasts its plan makes work for every vehicle that hears it — each one deserializes that trajectory and runs it against its own, twice a second, for every neighbour in range. Raise the fleet slider and broadcast crosses the processing budget while typed release does not.

The saving is in bytes and receiver work, not airtime: a packet costs about 150 µs of channel time almost regardless of how big it is.

What is real here, and what is simulated
  • Real: the road geometry and the surrounding traffic. NGSIM recordings of the Lankershim Boulevard and Peachtree Street signalised intersections, on OpenStreetMap lane graphs.
  • Real: every message decision, computed live by the paper’s selector as you drive.
  • Simulated, and has to be: your own car, which supplies the plan the rule needs a source to have, and the yielding, which follows the receiver controller from the paper. Your car also brakes for whatever is in front of it, so a run ends on a side impact or a head-on rather than on a shunt from traffic running with you.
  • Each surrounding vehicle holds the speed its own recording had at that point of the road, keeps a gap to the car in front, and brakes at a real rate for a conflict it has been told about. With the radio off they track their recordings exactly.
  • The recordings run about nine seconds. After that a vehicle carries on along the lane graph and leaves; new traffic enters at the edges at speeds taken from the recorded vehicles. Nothing is teleported back to where it came from, and a car that stands still for its whole recording is parked and stays parked.
  • The line you draw is not snapped to anything — go the wrong way up the boulevard if you like. The selector clears your trajectory whatever shape it has. The guarantee does rest on the map, but on the other side of the link: a receiver that left the lane graph the way you can would fall outside the reachable set your car computed for it.
  • The fleet gauge is an extrapolation. The share of neighbours handed a full plan is measured live from the decisions on screen; the rest asks what work a vehicle would be given if all of its neighbours behaved the way yours just did, against a stated budget of a thousand full-plan checks per second. No queueing or packet loss is modelled.
  • The carriageway is drawn wider than the lane it stands for, so traffic at true spacing does not read as a permanent near miss. Vehicle positions are untouched.