Skip to content

INTEGRATION_UNITY.md

Latest commit

 

History

History
59 lines (48 loc) · 2.96 KB

INTEGRATION_UNITY.md

File metadata and controls

59 lines (48 loc) · 2.96 KB
Raw

Integrating SimThink D with a Unity simulator

1. Process layout

  • The decider runs as a separate local process (simthinkd serve <decider>) on the same machine as the simulator.
  • The simulator is the HTTP client. One decision request in flight at a time per agent.
  • For remote-decider arms, point the same client at a relay URL or at mock_decider.py --delay-ms 250 --jitter-ms 30.

2. Asynchronous loop (C#, simplified)

// Fixed simulation step (e.g. 35 Hz) driven by an accumulator; rendering is independent.
class DecisionAgent {
    Task<Reply> pending; int sentTick; string lastAction = "ADVANCE";
    public void OnSimTick(int tick, SimState s) {
        if (pending != null && pending.IsCompleted) {
            var r = pending.Result;                       // apply at this tick
            lastAction = Valid(r) ? r.Choice : Fallback(); // unknown choice -> first action, logged invalid
            Log.Decision(sentTick, tick, r.RoundTripMs, lastAction);
            pending = null;
        }
        if (pending == null) {                            // request a new decision
            sentTick = tick;
            pending = client.PostAsync(BuildRequest(Perception.Sentence(s)));
        } else {
            Log.SkippedTick(tick);                        // decision still pending: this tick is skipped
        }
        Vehicle.Apply(lastAction);                        // the last action repeats while waiting
    }
}
  • Use HttpClient with a keep-alive connection and a 10 s timeout; measure round trip with Stopwatch.
  • Never block the simulation thread on the request. In synchronous (control) mode, pause the fixed step until the reply arrives.
  • Health check before a run: GET /health, compare weights_sha256 with the expected hash from the run config; abort on mismatch.

3. Arms to support

Arm Decides Waits for
T0 rule teacher (in C#) nothing
S rule teacher a real call to the decider (answer discarded)
D SimThink D its own call
J remote decider (URL) its own call
SJ rule teacher a replayed latency drawn from a file of measured latencies
INJ(k) any fixed extra delay k ms after each call
HOLD(k) decider holds the freshly applied action for k ticks before reading the next state

4. Logs

Per job (JSON, written atomically at the end): arm, scenario, seeds, tick rate, mode, url, expected hash, git commit, host, per-episode {seed, ticks, decisions, engine_ticks_skipped, decision_ms_p50, decision_ms_p90, invalid, outcome metrics, teacher_agreement} and optional gzip per-tick traces. These feed our analysis scripts unchanged.

5. Training data from the simulator

Run the rule teacher in synchronous mode on training seeds and write one row per decision in the format of simthinkd.toy and simthinkd.fit (request = exactly what the client would send; expected = the teacher's choice). Hand us the three .jsonl.gz files or train locally with simthinkd.train.