feat(simulation): add Worker AI polygon runtime and terrain navigation

2026-09-25 16:40:45 +03:00
parent a7c64e009d
commit f01bd39037
88 changed files with 9918 additions and 108 deletions
@@ -0,0 +1,87 @@
+"""Resident, pinned ADE20K surface provider. RGB only; no scene truth input."""
+
+import hashlib
+import json
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from pathlib import Path
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers import SegformerForSemanticSegmentation, SegformerImageProcessor
+
+# These are surface candidates, not permission to drive. The metric terrain
+# planner still checks step height, slope and the full rover footprint.
+SURFACES = (3, 6, 9, 11, 13, 29, 34, 46, 52, 91)
+FILES = {
+    "config.json": "ee7400840fdb1e5045f0b2eba78bf053df8e33a309c4acec31705a48c8cc5c00",
+    "preprocessor_config.json": "8039d1d210abaa7117ad78e58cdfd6141a2ec72c03dae891b3cd76737e422c6c",
+    "pytorch_model.bin": "187ca07bea003a5717c63d04ea90b07f33cd033c0ebf44b4b89fce5070d6c8f3",
+}
+
+
+def main():
+    for name, expected in FILES.items():
+        if hashlib.sha256((Path("/assets") / name).read_bytes()).hexdigest() != expected:
+            raise RuntimeError("SegFormer asset identity changed: " + name)
+    processor = SegformerImageProcessor.from_pretrained("/assets", local_files_only=True)
+    model = (
+        SegformerForSemanticSegmentation.from_pretrained("/assets", local_files_only=True)
+        .cuda()
+        .eval()
+    )
+    torch.set_num_threads(2)
+
+    def infer(rgb):
+        image = Image.fromarray(rgb).crop((100, 0, 700, 600))
+        tensor = processor(images=image, return_tensors="pt")["pixel_values"].cuda()
+        with torch.inference_mode():
+            logits = model(tensor).logits
+            logits = torch.nn.functional.interpolate(
+                logits, size=(512, 512), mode="bilinear", align_corners=False
+            )
+            confidence, labels = logits.softmax(1).max(1)
+        labels = labels[0].cpu().numpy().astype(np.uint8)
+        candidate = np.isin(labels, SURFACES) & (confidence[0].cpu().numpy() >= 0.55)
+        return labels.tobytes() + candidate.astype(np.uint8).tobytes()
+
+    infer(np.zeros((600, 800, 3), dtype=np.uint8))
+
+    class Handler(BaseHTTPRequestHandler):
+        def setup(self):
+            super().setup()
+            self.connection.settimeout(10)
+
+        def reply(self, status, body):
+            self.send_response(status)
+            self.send_header("Content-Type", "application/octet-stream")
+            self.send_header("Content-Length", str(len(body)))
+            self.end_headers()
+            self.wfile.write(body)
+
+        def do_GET(self):
+            self.reply(
+                200 if self.path == "/ready" else 404,
+                json.dumps({"model": "segformer-b2-ade150"}).encode(),
+            )
+
+        def do_POST(self):
+            if self.path != "/infer" or self.headers.get("Content-Length") != "1440000":
+                self.reply(400, b"Expected 800x600 raw RGB uint8")
+                return
+            try:
+                raw = self.rfile.read(1440000)
+                if len(raw) != 1440000:
+                    raise ValueError("Incomplete camera frame")
+                self.reply(200, infer(np.frombuffer(raw, np.uint8).reshape(600, 800, 3)))
+            except (TimeoutError, ValueError, RuntimeError):
+                self.reply(500, b"Surface inference failed")
+
+        def log_message(self, *_):
+            pass
+
+    HTTPServer(("0.0.0.0", 8010), Handler).serve_forever()
+
+
+if __name__ == "__main__":
+    main()