goal_gen_cases / scripts /refine_pose_physics.py
yianW's picture
Initial upload: 8 manipulation cases (place_scene + physics refine + assets + genesis runtime patch)
80f0adb verified
"""
Stage 6: Physics-based pose refinement.
Given an estimated pose from Stage 5, sample poses around it and validate
using physics simulation. A pose is valid if:
1. No collision with container at initial placement
2. Position doesn't change much after physics settling
3. Rotation doesn't change much after physics settling
"""
import json
import os
from dataclasses import dataclass
from typing import List, Optional, Tuple
import numpy as np
import torch
from scipy.spatial.transform import Rotation as R
import genesis as gs
import genesis.utils.geom as gu
import genesis_patch # call genesis_patch.apply() AFTER gs.init()
@dataclass
class PhysicsRefineArgs:
"""Configuration for physics-based pose refinement."""
output_dir: str
object_mesh_path: str
container_mesh_path: str
table_path: str = "../video_gen_assets/table.glb"
# Scene layout (must match place_scene.py defaults)
table_pos: Tuple[float, float, float] = (0.0, 0.0, 0.485403)
table_euler: Tuple[float, float, float] = (90.0, 0.0, 90.0)
table_height: float = 0.7451
container_euler: Tuple[float, float, float] = (90.0, 0.0, 0.0)
container_pos: Optional[Tuple[float, float, float]] = None # Override auto-z placement
container_scale: float = 1.0 # Scale factor for container mesh
object_scale: float = 1.0 # Scale factor for target object mesh
# Sampling parameters
num_samples: int = 64 # Number of poses to sample (use power of 2 for parallel sim)
xy_jitter: float = 0.02 # meters
z_jitter: float = 0.01 # meters
xyz_jitter: Tuple[float, float, float] = None # per-axis jitter (overrides xy/z)
rot_jitter_deg: Tuple[float, float, float] = (5.0, 5.0, 15.0) # degrees
# Physics parameters
physics_steps: int = 300 # Number of simulation steps
dt: float = 0.01 # Simulation timestep
# Validation thresholds
max_pos_drift: float = 0.03 # meters - max allowed position change
max_rot_drift_deg: float = 15.0 # degrees - max allowed rotation change
# Use parallel simulation
use_parallel: bool = True
# Wall-mounted scene (e.g., hanger_hooks)
wall_mode: bool = False
wall_x: float = -0.5
wall_height: float = 1.2
# CoACD decomposition parameters
coacd_threshold: float = 0.01
coacd_resolution: int = 80
max_convex_hull: int = 20
# Multi-round sampling
num_rounds: int = 1 # Number of rounds (each round = num_samples parallel envs)
# Render initial pose for debugging
render_initial: bool = False
def jitter_pos(pos: np.ndarray, xy_jitter: float = 0.02, z_jitter: float = 0.01,
xyz_jitter: Tuple[float, float, float] = None) -> np.ndarray:
"""Add small uniform noise to position. If xyz_jitter is given, use per-axis values."""
pos = np.array(pos, dtype=np.float32).copy()
if xyz_jitter is not None:
for i in range(3):
pos[i] += np.random.uniform(-xyz_jitter[i], xyz_jitter[i])
else:
pos[:2] += np.random.uniform(-xy_jitter, xy_jitter, size=2)
pos[2] += np.random.uniform(-z_jitter, z_jitter)
return pos
def jitter_rot(euler_deg: np.ndarray, jitter_deg: Tuple[float, float, float] = (5, 5, 15)) -> np.ndarray:
"""Add small uniform noise to each Euler angle (degrees)."""
euler_deg = np.array(euler_deg, dtype=np.float32).copy()
jitter = np.array(jitter_deg)
euler_deg += np.random.uniform(-jitter, jitter)
return euler_deg
def rotvec_to_euler_deg(rotvec: np.ndarray) -> np.ndarray:
"""Convert rotation vector to Euler angles in degrees."""
rot = R.from_rotvec(rotvec)
return rot.as_euler('xyz', degrees=True)
def euler_deg_to_quat_wxyz(euler_deg: np.ndarray) -> np.ndarray:
"""Convert Euler angles (degrees) to quaternion (w, x, y, z)."""
rot = R.from_euler('xyz', euler_deg, degrees=True)
quat_xyzw = rot.as_quat()
return np.array([quat_xyzw[3], quat_xyzw[0], quat_xyzw[1], quat_xyzw[2]])
def compute_rotation_diff_deg(quat1_wxyz: np.ndarray, quat2_wxyz: np.ndarray) -> float:
"""Compute rotation difference in degrees between two quaternions."""
# Convert to scipy format (xyzw)
q1_xyzw = np.array([quat1_wxyz[1], quat1_wxyz[2], quat1_wxyz[3], quat1_wxyz[0]])
q2_xyzw = np.array([quat2_wxyz[1], quat2_wxyz[2], quat2_wxyz[3], quat2_wxyz[0]])
r1 = R.from_quat(q1_xyzw)
r2 = R.from_quat(q2_xyzw)
# Compute relative rotation
r_diff = r1.inv() * r2
angle_rad = r_diff.magnitude()
return np.degrees(angle_rad)
def _load_estimated_pose(args: PhysicsRefineArgs):
"""Load estimated pose from Stage 5 results."""
pose_path = os.path.join(args.output_dir, "pose_estimation", "best_pose.json")
with open(pose_path, "r") as f:
pose_data = json.load(f)
if "tvec_w_icp" in pose_data:
tvec_w = np.array(pose_data["tvec_w_icp"])
rvec_w = np.array(pose_data["rvec_w_icp"])
print("Using ICP-refined pose as starting point")
else:
tvec_w = np.array(pose_data["tvec_w"])
rvec_w = np.array(pose_data["rvec_w"])
print("Using 3D-3D registration pose as starting point")
print(f"Starting position: {tvec_w}")
print(f"Starting rotation (rotvec): {rvec_w}")
euler_deg = rotvec_to_euler_deg(rvec_w)
return tvec_w, rvec_w, euler_deg
def _build_scene(args: PhysicsRefineArgs, tvec_w, euler_deg):
"""Build Genesis scene for physics refinement (table or wall mode)."""
import trimesh
from trimesh.transformations import euler_matrix
gs.init(seed=0, precision="32", logging_level="warning")
genesis_patch.apply() # adds RigidEntity.detect_collision_parallel
# Add renderer if rendering is requested (Rasterizer handles collision meshes)
renderer_kwargs = {}
if args.render_initial:
renderer_kwargs["renderer"] = gs.renderers.Rasterizer()
scene = gs.Scene(
sim_options=gs.options.SimOptions(dt=args.dt, substeps=5),
show_viewer=False,
rigid_options=gs.options.RigidOptions(
enable_collision=True,
enable_self_collision=False,
),
**renderer_kwargs,
)
mat_convex = gs.materials.Rigid()
mat_target = gs.materials.Rigid(friction=3.0)
wall = None
coacd_opts = gs.options.CoacdOptions(
threshold=args.coacd_threshold,
preprocess_resolution=args.coacd_resolution,
decimate=True,
max_convex_hull=args.max_convex_hull,
)
if args.wall_mode:
# Wall-mounted scene
from place_scene import add_wall, DEFAULT_WALL_TEXTURE
c_euler = list(args.container_euler)
# Add wall as a collision plane (box)
wall_thickness = 0.05
wall = scene.add_entity(
material=gs.materials.Rigid(sdf_min_res=16, sdf_max_res=16),
morph=gs.morphs.Box(
pos=[args.wall_x - wall_thickness / 2, 0.0, 1.5],
size=[wall_thickness, 3.0, 3.0],
collision=True,
fixed=True,
),
)
# Place container flush against wall
crx, cry, crz = np.deg2rad(c_euler)
cmesh = trimesh.load(args.container_mesh_path, force="mesh")
cmesh.apply_transform(euler_matrix(crx, cry, crz, "sxyz"))
min_x = cmesh.bounds[0][0]
container_x = args.wall_x - min_x
container_pos = [container_x, 0.0, args.wall_height]
print(f"Container placed on wall at pos={container_pos}")
container = scene.add_entity(
material=mat_convex,
morph=gs.morphs.Mesh(
file=args.container_mesh_path,
pos=container_pos,
euler=c_euler,
scale=args.container_scale,
collision=True,
fixed=True,
convexify=True,
merge_submeshes_for_collision=False,
group_by_material=False,
coacd_options=coacd_opts,
),
vis_mode="collision",
)
else:
# Table-mounted scene
mat_sdf_lo = gs.materials.Rigid(sdf_min_res=16, sdf_max_res=16)
c_euler = list(args.container_euler)
scene.add_entity(
material=mat_sdf_lo,
morph=gs.morphs.Mesh(
file=args.table_path,
pos=list(args.table_pos),
euler=list(args.table_euler),
collision=True,
fixed=True,
convexify=True,
),
)
if args.container_pos is not None:
container_pos = list(args.container_pos)
print(f"Container placed at pos={container_pos} (override)")
else:
crx, cry, crz = np.deg2rad(c_euler)
cmesh = trimesh.load(args.container_mesh_path, force="mesh")
cmesh.apply_transform(euler_matrix(crx, cry, crz, "sxyz"))
container_z = args.table_height - cmesh.bounds[0][2]
container_pos = [0.0, 0.0, container_z]
print(f"Container auto-placed at z={container_z:.4f}")
container = scene.add_entity(
material=mat_convex,
morph=gs.morphs.Mesh(
file=args.container_mesh_path,
pos=container_pos,
euler=c_euler,
scale=args.container_scale,
collision=True,
fixed=True,
convexify=True,
merge_submeshes_for_collision=False,
group_by_material=False,
coacd_options=coacd_opts,
),
vis_mode="collision",
)
# Add target object
target_mesh_kwargs = {
"file": args.object_mesh_path,
"pos": list(tvec_w),
"euler": list(euler_deg),
"collision": True,
"fixed": False,
"coacd_options": coacd_opts,
}
if args.object_scale != 1.0:
target_mesh_kwargs["scale"] = args.object_scale
target = scene.add_entity(
material=mat_target,
morph=gs.morphs.Mesh(**target_mesh_kwargs),
vis_mode="collision",
)
for link in target.links:
link._inertial_mass = 0.05
# Add camera if rendering is requested (must be before scene.build)
cam = None
if args.render_initial:
cam_pose_path = os.path.join(args.output_dir, "cam_pose.npy")
if os.path.exists(cam_pose_path):
cam_pose_mat = np.load(cam_pose_path)
cam_pos = cam_pose_mat[:3, 3]
forward = cam_pose_mat[:3, 2]
lookat = cam_pos + forward * 1.0
cam = scene.add_camera(
pos=tuple(cam_pos),
lookat=tuple(lookat),
res=(1200, 896),
fov=50.0,
GUI=False,
)
else:
print("WARNING: cam_pose.npy not found, skipping render")
return scene, container, target, wall, cam
def _render_initial_pose(args: PhysicsRefineArgs, cam):
"""Render the initial pose with collision meshes for debugging."""
import imageio
if cam is None:
return
img = cam.render()[0]
out_dir = os.path.join(args.output_dir, "physics_refinement")
os.makedirs(out_dir, exist_ok=True)
out_path = os.path.join(out_dir, "initial_pose_collision_mesh.png")
imageio.imwrite(out_path, img)
print(f"Rendered initial pose with collision meshes: {out_path}")
def refine_pose_physics_sequential(args: PhysicsRefineArgs) -> List[dict]:
"""
Refine pose using sequential physics simulation.
Simpler but slower than parallel version.
"""
print("=" * 60)
print("Physics-based Pose Refinement (Sequential)")
print("=" * 60)
tvec_w, rvec_w, euler_deg = _load_estimated_pose(args)
scene, container, target, wall_ent, _ = _build_scene(args, tvec_w, euler_deg)
scene.build()
# Sample and validate poses
valid_poses = []
print(f"\nSampling {args.num_samples} poses around estimated position...")
for i in range(args.num_samples):
# Jitter pose
sampled_pos = jitter_pos(tvec_w, args.xy_jitter, args.z_jitter, xyz_jitter=args.xyz_jitter)
sampled_euler = jitter_rot(euler_deg, args.rot_jitter_deg)
sampled_quat = euler_deg_to_quat_wxyz(sampled_euler)
# Reset scene and set pose
scene.reset()
target.set_pos(torch.tensor(sampled_pos, dtype=torch.float32))
target.set_quat(torch.tensor(sampled_quat, dtype=torch.float32))
# Record initial pose
init_pos = target.get_pos().cpu().numpy()
init_quat = target.get_quat().cpu().numpy()
# Run physics simulation
for _ in range(args.physics_steps):
scene.step()
# Get final pose
final_pos = target.get_pos().cpu().numpy()
final_quat = target.get_quat().cpu().numpy()
# Compute drift
pos_drift = np.linalg.norm(final_pos - init_pos)
rot_drift = compute_rotation_diff_deg(init_quat, final_quat)
# Check validity
is_valid = (pos_drift < args.max_pos_drift) and (rot_drift < args.max_rot_drift_deg)
if is_valid:
valid_poses.append({
"pos": final_pos.tolist(),
"quat_wxyz": final_quat.tolist(),
"euler_deg": rotvec_to_euler_deg(R.from_quat(
[final_quat[1], final_quat[2], final_quat[3], final_quat[0]]
).as_rotvec()).tolist(),
"pos_drift": float(pos_drift),
"rot_drift_deg": float(rot_drift),
"sample_idx": i,
})
print(f" Sample {i}: VALID (pos_drift={pos_drift:.4f}m, rot_drift={rot_drift:.2f}°)")
else:
if i < 10: # Only print first few invalid ones
print(f" Sample {i}: invalid (pos_drift={pos_drift:.4f}m, rot_drift={rot_drift:.2f}°)")
print(f"\nFound {len(valid_poses)} valid poses out of {args.num_samples} samples")
return valid_poses
def refine_pose_physics_parallel(args: PhysicsRefineArgs) -> List[dict]:
"""
Refine pose using parallel physics simulation with batched environments.
Supports multiple rounds of sampling.
"""
print("=" * 60)
print("Physics-based Pose Refinement (Parallel)")
print("=" * 60)
tvec_w, rvec_w, euler_deg = _load_estimated_pose(args)
n_envs = args.num_samples
scene, container, target, wall_ent, cam = _build_scene(args, tvec_w, euler_deg)
# Build with parallel environments
scene.build(n_envs=n_envs)
# Render initial pose if requested (after build, before sim rounds)
if args.render_initial:
_render_initial_pose(args, cam)
all_valid_poses = []
total_samples = 0
total_collisions = 0
for round_idx in range(args.num_rounds):
if args.num_rounds > 1:
print(f"\n--- Round {round_idx + 1}/{args.num_rounds} ---")
# Pre-generate sampled poses for this round
sampled_positions = np.zeros((n_envs, 3), dtype=np.float32)
sampled_quats = np.zeros((n_envs, 4), dtype=np.float32)
for i in range(n_envs):
if round_idx == 0 and i == 0:
# First sample of first round: original pose (no jitter)
sampled_pos = np.array(tvec_w, dtype=np.float32)
sampled_quat = euler_deg_to_quat_wxyz(euler_deg)
else:
sampled_pos = jitter_pos(tvec_w, args.xy_jitter, args.z_jitter, xyz_jitter=args.xyz_jitter)
sampled_euler = jitter_rot(euler_deg, args.rot_jitter_deg)
sampled_quat = euler_deg_to_quat_wxyz(sampled_euler)
sampled_positions[i] = sampled_pos
sampled_quats[i] = sampled_quat
print(f"\nRunning {n_envs} parallel simulations...")
# Reset scene state for new round
scene.reset()
# Set initial poses for all environments
target.set_pos(torch.tensor(sampled_positions, dtype=torch.float32, device="cuda"))
target.set_quat(torch.tensor(sampled_quats, dtype=torch.float32, device="cuda"))
# Check initial collision before stepping (no state change)
collided = target.detect_collision_parallel(with_entity=container)
if wall_ent is not None:
collided_wall = target.detect_collision_parallel(with_entity=wall_ent)
collided = collided | collided_wall
n_colliding = int(collided.sum())
total_collisions += n_colliding
total_samples += n_envs
print(f" Initial collision: {n_colliding}/{n_envs} envs")
# Record initial poses
init_pos = target.get_pos().cpu().numpy() # (n_envs, 3)
init_quat = target.get_quat().cpu().numpy() # (n_envs, 4)
# Run physics simulation
for step in range(args.physics_steps):
scene.step()
if (step + 1) % 30 == 0:
print(f" Step {step + 1}/{args.physics_steps}")
# Get final poses
final_pos = target.get_pos().cpu().numpy() # (n_envs, 3)
final_quat = target.get_quat().cpu().numpy() # (n_envs, 4)
# Validate all poses
round_valid = 0
for i in range(n_envs):
if collided[i]:
continue
pos_drift = np.linalg.norm(final_pos[i] - init_pos[i])
rot_drift = compute_rotation_diff_deg(init_quat[i], final_quat[i])
is_valid = (pos_drift < args.max_pos_drift) and (rot_drift < args.max_rot_drift_deg)
if is_valid:
quat_xyzw = [final_quat[i][1], final_quat[i][2], final_quat[i][3], final_quat[i][0]]
final_euler = R.from_quat(quat_xyzw).as_euler('xyz', degrees=True)
init_quat_xyzw = [init_quat[i][1], init_quat[i][2], init_quat[i][3], init_quat[i][0]]
init_euler = R.from_quat(init_quat_xyzw).as_euler('xyz', degrees=True)
all_valid_poses.append({
"pos": final_pos[i].tolist(),
"quat_wxyz": final_quat[i].tolist(),
"euler_deg": final_euler.tolist(),
"init_pos": init_pos[i].tolist(),
"init_quat_wxyz": init_quat[i].tolist(),
"init_euler_deg": init_euler.tolist(),
"pos_drift": float(pos_drift),
"rot_drift_deg": float(rot_drift),
"sample_idx": round_idx * n_envs + i,
"round": round_idx,
})
round_valid += 1
print(f" Round {round_idx + 1}: {round_valid} valid poses")
print(f"\nTotal: {len(all_valid_poses)} valid poses out of {total_samples} samples "
f"({total_collisions} initial collisions)")
# Sort by stability (least drift)
all_valid_poses.sort(key=lambda x: x["pos_drift"] + x["rot_drift_deg"] / 100)
return all_valid_poses
def refine_pose_physics(args: PhysicsRefineArgs) -> dict:
"""
Main entry point for physics-based pose refinement.
Returns dict with best refined pose and all valid poses.
"""
if args.use_parallel:
valid_poses = refine_pose_physics_parallel(args)
else:
valid_poses = refine_pose_physics_sequential(args)
# Save results
refine_dir = os.path.join(args.output_dir, "physics_refinement")
os.makedirs(refine_dir, exist_ok=True)
total_samples = args.num_samples * args.num_rounds
results = {
"num_samples": total_samples,
"num_rounds": args.num_rounds,
"num_valid": len(valid_poses),
"valid_poses": valid_poses,
"best_pose": valid_poses[0] if valid_poses else None,
}
results_path = os.path.join(refine_dir, "refined_poses.json")
with open(results_path, "w") as f:
json.dump(results, f, indent=2)
print(f"\nSaved results to: {results_path}")
if valid_poses:
best = valid_poses[0]
print(f"\nBest refined pose:")
print(f" Position: {best['pos']}")
print(f" Rotation (euler): {best['euler_deg']}")
print(f" Stability: pos_drift={best['pos_drift']:.4f}m, rot_drift={best['rot_drift_deg']:.2f}°")
return results
def main():
import argparse
parser = argparse.ArgumentParser(description="Physics-based pose refinement")
parser.add_argument("--output_dir", type=str, default="./outputs/mug_tree_goal",
help="Output directory with pipeline outputs")
parser.add_argument("--object_mesh_path", type=str, default="./assets/mug_0.glb",
help="Path to target object mesh")
parser.add_argument("--container_mesh_path", type=str,
default="./assets/metal_mug_holder/metal_mug_holder.obj",
help="Path to container mesh")
parser.add_argument("--table_path", type=str,
default="../video_gen_assets/table.glb",
help="Path to table mesh")
parser.add_argument("--num_samples", type=int, default=64,
help="Number of poses to sample")
parser.add_argument("--sequential", action="store_true",
help="Use sequential simulation instead of parallel")
parser.add_argument("--container_euler", type=float, nargs=3, default=[90.0, 0.0, 0.0],
help="Container euler angles in degrees (default: 90 0 0)")
parser.add_argument("--container_pos", type=float, nargs=3, default=None,
help="Override container position [x, y, z] (skip auto-z computation)")
parser.add_argument("--xy_jitter", type=float, default=0.02,
help="XY jitter range in meters (default: 0.02)")
parser.add_argument("--z_jitter", type=float, default=0.01,
help="Z jitter range in meters (default: 0.01)")
parser.add_argument("--xyz_jitter", type=float, nargs=3, default=None,
help="Per-axis jitter [x, y, z] in meters (overrides xy/z_jitter)")
parser.add_argument("--rot_jitter", type=float, nargs=3, default=[5.0, 5.0, 15.0],
help="Rotation jitter in degrees per axis (default: 5 5 15)")
parser.add_argument("--wall", action="store_true",
help="Use wall-mounted scene instead of table")
parser.add_argument("--wall_x", type=float, default=-0.5,
help="X position of wall plane (default: -0.5)")
parser.add_argument("--wall_height", type=float, default=1.2,
help="Z height to mount container on wall (default: 1.2)")
parser.add_argument("--container_scale", type=float, default=1.0,
help="Scale factor for container mesh (default: 1.0)")
parser.add_argument("--object_scale", type=float, default=1.0,
help="Scale factor for target object mesh (default: 1.0)")
parser.add_argument("--coacd_threshold", type=float, default=0.01,
help="CoACD threshold (default: 0.01)")
parser.add_argument("--coacd_resolution", type=int, default=80,
help="CoACD preprocess resolution (default: 80)")
parser.add_argument("--max_convex_hull", type=int, default=20,
help="CoACD max convex hulls (default: 20)")
parser.add_argument("--num_rounds", type=int, default=1,
help="Number of sampling rounds (default: 1)")
parser.add_argument("--max_pos_drift", type=float, default=0.03,
help="Max position drift for valid pose (meters, default: 0.03)")
parser.add_argument("--max_rot_drift", type=float, default=15.0,
help="Max rotation drift for valid pose (degrees, default: 15.0)")
parser.add_argument("--render_initial", action="store_true",
help="Render initial pose with collision meshes for debugging")
cli_args = parser.parse_args()
args = PhysicsRefineArgs(
output_dir=cli_args.output_dir,
object_mesh_path=cli_args.object_mesh_path,
container_mesh_path=cli_args.container_mesh_path,
table_path=cli_args.table_path,
num_samples=cli_args.num_samples,
use_parallel=not cli_args.sequential,
container_euler=tuple(cli_args.container_euler),
container_pos=tuple(cli_args.container_pos) if cli_args.container_pos else None,
xy_jitter=cli_args.xy_jitter,
z_jitter=cli_args.z_jitter,
rot_jitter_deg=tuple(cli_args.rot_jitter),
wall_mode=cli_args.wall,
wall_x=cli_args.wall_x,
wall_height=cli_args.wall_height,
coacd_threshold=cli_args.coacd_threshold,
coacd_resolution=cli_args.coacd_resolution,
max_convex_hull=cli_args.max_convex_hull,
num_rounds=cli_args.num_rounds,
render_initial=cli_args.render_initial,
object_scale=cli_args.object_scale,
max_pos_drift=cli_args.max_pos_drift,
max_rot_drift_deg=cli_args.max_rot_drift,
)
refine_pose_physics(args)
if __name__ == "__main__":
main()