update

embodied_gen/data/convex_decomposer.py

@@ -19,6 +19,7 @@ import multiprocessing as mp
 import os
 
 import coacd
+import numpy as np
 import trimesh
 
 logger = logging.getLogger(__name__)
@@ -31,7 +32,11 @@ __all__ = [
 
 
 def decompose_convex_coacd(
-    filename: str, outfile: str, params: dict, verbose: bool = False
+    filename: str,
+    outfile: str,
+    params: dict,
+    verbose: bool = False,
+    auto_scale: bool = True,
 ) -> None:
     coacd.set_log_level("info" if verbose else "warn")
 
@@ -40,6 +45,14 @@ def decompose_convex_coacd(
 
     result = coacd.run_coacd(mesh, **params)
     combined = sum([trimesh.Trimesh(*m) for m in result])
+
+    # Compute collision_scale because convex decomposition usually makes the mesh larger.
+    if auto_scale:
+        convex_mesh_shape = np.ptp(combined.vertices, axis=0)
+        visual_mesh_shape = np.ptp(mesh.vertices, axis=0)
+        rescale = visual_mesh_shape / convex_mesh_shape
+        combined.vertices *= rescale
+
     combined.export(outfile)
 
 
@@ -57,6 +70,7 @@ def decompose_convex_mesh(
     pca: bool = False,
     merge: bool = True,
     seed: int = 0,
+    auto_scale: bool = True,
     verbose: bool = False,
 ) -> str:
     """Decompose a mesh into convex parts using the CoACD algorithm."""
@@ -81,7 +95,7 @@ def decompose_convex_mesh(
     )
 
     try:
-        decompose_convex_coacd(filename, outfile, params, verbose)
+        decompose_convex_coacd(filename, outfile, params, verbose, auto_scale)
         if os.path.exists(outfile):
             return outfile
     except Exception as e:
@@ -91,7 +105,9 @@ def decompose_convex_mesh(
     if preprocess_mode != "on":
         try:
             params["preprocess_mode"] = "on"
-            decompose_convex_coacd(filename, outfile, params, verbose)
+            decompose_convex_coacd(
+                filename, outfile, params, verbose, auto_scale
+            )
             if os.path.exists(outfile):
                 return outfile
         except Exception as e:
@@ -118,6 +134,7 @@ def decompose_convex_mp(
     merge: bool = True,
     seed: int = 0,
     verbose: bool = False,
+    auto_scale: bool = True,
 ) -> str:
     """Decompose a mesh into convex parts using the CoACD algorithm in a separate process.
 
@@ -140,7 +157,7 @@ def decompose_convex_mp(
     ctx = mp.get_context("spawn")
     p = ctx.Process(
         target=decompose_convex_coacd,
-        args=(filename, outfile, params, verbose),
+        args=(filename, outfile, params, verbose, auto_scale),
     )
     p.start()
     p.join()
@@ -151,7 +168,7 @@ def decompose_convex_mp(
         params["preprocess_mode"] = "on"
         p = ctx.Process(
             target=decompose_convex_coacd,
-            args=(filename, outfile, params, verbose),
+            args=(filename, outfile, params, verbose, auto_scale),
         )
         p.start()
         p.join()

embodied_gen/models/text_model.py

@@ -136,7 +136,7 @@ def build_text2img_ip_pipeline(
     pipe = pipe.to(device)
     pipe.image_encoder = pipe.image_encoder.to(device)
     # pipe.enable_model_cpu_offload()
-    # pipe.enable_xformers_memory_efficient_attention()
+    # # pipe.enable_xformers_memory_efficient_attention()
     # pipe.enable_vae_slicing()
 
     return pipe

embodied_gen/scripts/gen_layout.py

@@ -137,7 +137,7 @@ def entrypoint() -> None:
         sim_cli(
             layout_path=layout_path,
             output_dir=output_root,
-            robot_name="franka" if args.insert_robot else None,
+            insert_robot=args.insert_robot,
         )
 
         torch.cuda.empty_cache()

embodied_gen/scripts/simulate_sapien.py

@@ -156,14 +156,15 @@ def entrypoint(**kwargs):
         bg_images[camera.name] = result.rgb[..., ::-1]
 
     video_frames = []
-    for camera in scene_manager.cameras:
+    for idx, camera in enumerate(scene_manager.cameras):
         # Scene rendering
-        for step in range(image_cnt):
-            rgba = alpha_blend_rgba(
-                frames[camera.name][step]["Foreground"],
-                bg_images[camera.name],
-            )
-            video_frames.append(np.array(rgba))
+        if idx == 0:
+            for step in range(image_cnt):
+                rgba = alpha_blend_rgba(
+                    frames[camera.name][step]["Foreground"],
+                    bg_images[camera.name],
+                )
+                video_frames.append(np.array(rgba))
 
         # Grasp rendering
         for node in actions:

embodied_gen/utils/simulation.py

@@ -73,8 +73,17 @@ def load_actor_from_urdf(
     root = tree.getroot()
     node_name = root.get("name")
     file_dir = os.path.dirname(file_path)
-    visual_file = root.find('.//visual/geometry/mesh').get("filename")
-    collision_file = root.find('.//collision/geometry/mesh').get("filename")
+
+    visual_mesh = root.find('.//visual/geometry/mesh')
+    visual_file = visual_mesh.get("filename")
+    visual_scale = visual_mesh.get("scale", "1.0 1.0 1.0")
+    visual_scale = np.array([float(x) for x in visual_scale.split()])
+
+    collision_mesh = root.find('.//collision/geometry/mesh')
+    collision_file = collision_mesh.get("filename")
+    collision_scale = collision_mesh.get("scale", "1.0 1.0 1.0")
+    collision_scale = np.array([float(x) for x in collision_scale.split()])
+
     visual_file = os.path.join(file_dir, visual_file)
     collision_file = os.path.join(file_dir, collision_file)
     static_fric = root.find('.//collision/gazebo/mu1').text
@@ -90,15 +99,16 @@ def load_actor_from_urdf(
     body_type = "static" if use_static else "dynamic"
     builder.set_physx_body_type(body_type)
     builder.add_multiple_convex_collisions_from_file(
-        collision_file if body_type == "dynamic" else visual_file,
+        collision_file,
         material=material,
+        scale=collision_scale,
         # decomposition="coacd",
         # decomposition_params=dict(
         #     threshold=0.05, max_convex_hull=64, verbose=False
         # ),
     )
 
-    builder.add_visual_from_file(visual_file)
+    builder.add_visual_from_file(visual_file, scale=visual_scale)
     builder.set_initial_pose(pose)
     if isinstance(scene, ManiSkillScene) and env_idx is not None:
         builder.set_scene_idxs([env_idx])

embodied_gen/validators/urdf_convertor.py

@@ -134,8 +134,9 @@ class URDFGenerator(object):
             Estimate the vertical projection of their real length based on its pose.
             For example:
               - A pen standing upright in the first image (aligned with the image's vertical axis)
-              full body visible in the first image: → vertical height ≈ 0.14-0.20 m
-              - A pen lying flat in the first image (showing thickness or as a dot) → vertical height ≈ 0.018-0.025 m
+                full body visible in the first image: → vertical height ≈ 0.14-0.20 m
+              - A pen lying flat in the first image or either the tip or the tail is facing the image
+                (showing thickness or as a circle) → vertical height ≈ 0.018-0.025 m
               - Tilted pen in the first image (e.g., ~45° angle): vertical height ≈ 0.07-0.12 m
             - Use the rest views to help determine the object's 3D pose and orientation.
             Assume the object is in real-world scale and estimate the approximate vertical height