From c62fba40b26fd2540fcf0e4ecbb75447aa9cb1a9 Mon Sep 17 00:00:00 2001
From: Paul Huliganga <paje0101@gmail.com>
Date: Thu, 14 May 2026 15:32:10 -0400
Subject: [PATCH] fix(agent): explicit hit backstop in StuckTermination + eval
 diagnostics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

multitrack_runner.py: adds Python-side hit check as a zero-latency backstop
— gym_donkeycar can delay hit!=none termination by one frame; this fires
on the same step and records stuck_reason for diagnostics.

eval_on_track.py: logs hit value and stuck_reason at episode end; calls
exit_scene after eval so the sim returns to main menu (next gym.make() can
switch scenes); removes unused SPEED_SCALE constant.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 agent/eval_on_track.py     | 17 ++++++++++++++---
 agent/multitrack_runner.py | 10 ++++++++++
 2 files changed, 24 insertions(+), 3 deletions(-)

diff --git a/agent/eval_on_track.py b/agent/eval_on_track.py
index 63065bf..e1e3d5d 100644
--- a/agent/eval_on_track.py
+++ b/agent/eval_on_track.py
@@ -22,13 +22,12 @@ from donkeycar_sb3_runner import ThrottleClampWrapper
 from multitrack_runner import StuckTerminationWrapper
 
 THROTTLE_MIN = 0.2
-SPEED_SCALE  = 0.1
 
 def make_env(env_id):
     raw = gym.make(env_id)
     env = ThrottleClampWrapper(raw, throttle_min=THROTTLE_MIN)
     env = StuckTerminationWrapper(env, stuck_steps=80, min_displacement=0.5)
-    env = SpeedRewardWrapper(env, speed_scale=SPEED_SCALE)
+    env = SpeedRewardWrapper(env)
     return env
 
 def main():
@@ -82,9 +81,15 @@ def main():
             if eff < 0.3 and steps >= 500:
                 note = f'  ⚠️  SHUTTLE EXPLOIT? macro_eff={eff:.2f}'
 
+        raw_info_final = info[0] if isinstance(info, (list, tuple)) else info
+        hit_val = raw_info_final.get('hit', '?') if isinstance(raw_info_final, dict) else '?'
+        stuck_reason = raw_info_final.get('stuck_reason', '') if isinstance(raw_info_final, dict) else ''
+        term_note = f'  hit={hit_val}' if hit_val not in ('?', 'none', None) else ''
+        if stuck_reason:
+            term_note += f'  reason={stuck_reason}'
         status = '✅ RAN FULL EVAL' if steps >= args.max_steps else '❌ CRASHED'
         print(f'  ep{ep+1}: {total_reward:.1f} reward / {steps} steps '
-              f'({total_reward/max(steps,1):.2f}/step)  {status}{note}')
+              f'({total_reward/max(steps,1):.2f}/step)  {status}{note}{term_note}')
         all_rewards.append(total_reward)
         all_steps.append(steps)
         time.sleep(0.5)
@@ -93,6 +98,12 @@ def main():
     print(f'  Mean steps  : {np.mean(all_steps):.0f}')
     print(f'  {"✅ DRIVES" if np.mean(all_steps) > 500 else "❌ CRASHES"}')
 
+    # Send exit_scene so sim returns to main menu — next gym.make() can switch scenes
+    try:
+        raw_env.unwrapped.viewer.exit_scene()
+        time.sleep(3)
+    except Exception:
+        pass
     env.close()
     time.sleep(2)
 
diff --git a/agent/multitrack_runner.py b/agent/multitrack_runner.py
index 0096da5..30b47cd 100644
--- a/agent/multitrack_runner.py
+++ b/agent/multitrack_runner.py
@@ -211,6 +211,16 @@ class StuckTerminationWrapper(gym.Wrapper):
             except (TypeError, ValueError):
                 pass
 
+        # Explicit hit check: gym_donkeycar sets done=True for hit!="none" but timing
+        # gaps (script execution order, 1-frame delay) can let it slip through.
+        # This is a zero-latency Python-side backstop — fires on the same step as the hit.
+        if not terminated:
+            hit = info.get('hit', 'none')
+            if hit and hit != 'none':
+                terminated = True
+                info['stuck_termination'] = True
+                info['stuck_reason'] = f'hit_{hit}'
+
         # Speed-based stuck detection: catches car pinned against a barrier.
         # A car pressed against a wall has speed≈0 even while sliding laterally
         # (accumulating displacement that resets the position-based timer above).