unitree_rl_gym/legged_gym/utils/isaacgym_utils.py

585 lines
19 KiB
Python

import os
import numpy as np
import random
import torch
def set_seed(seed):
if seed == -1:
seed = np.random.randint(0, 10000)
print("Setting seed: {}".format(seed))
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def to_torch(x, dtype=torch.float, device='cuda:0', requires_grad=False):
return torch.tensor(x,
dtype=dtype,
device=device,
requires_grad=requires_grad)
@torch.jit.script
def quat_mul(a, b):
assert a.shape == b.shape
shape = a.shape
a = a.reshape(-1, 4)
b = b.reshape(-1, 4)
x1, y1, z1, w1 = a[:, 0], a[:, 1], a[:, 2], a[:, 3]
x2, y2, z2, w2 = b[:, 0], b[:, 1], b[:, 2], b[:, 3]
ww = (z1 + x1) * (x2 + y2)
yy = (w1 - y1) * (w2 + z2)
zz = (w1 + y1) * (w2 - z2)
xx = ww + yy + zz
qq = 0.5 * (xx + (z1 - x1) * (x2 - y2))
w = qq - ww + (z1 - y1) * (y2 - z2)
x = qq - xx + (x1 + w1) * (x2 + w2)
y = qq - yy + (w1 - x1) * (y2 + z2)
z = qq - zz + (z1 + y1) * (w2 - x2)
quat = torch.stack([x, y, z, w], dim=-1).view(shape)
return quat
@torch.jit.script
def normalize(x, eps: float = 1e-9):
return x / x.norm(p=2, dim=-1).clamp(min=eps, max=None).unsqueeze(-1)
@torch.jit.script
def quat_apply(a, b):
shape = b.shape
a = a.reshape(-1, 4)
b = b.reshape(-1, 3)
xyz = a[:, :3]
t = xyz.cross(b, dim=-1) * 2
return (b + a[:, 3:] * t + xyz.cross(t, dim=-1)).view(shape)
@torch.jit.script
def quat_rotate(q, v):
shape = q.shape
q_w = q[:, -1]
q_vec = q[:, :3]
a = v * (2.0 * q_w ** 2 - 1.0).unsqueeze(-1)
b = torch.cross(q_vec, v, dim=-1) * q_w.unsqueeze(-1) * 2.0
c = q_vec * \
torch.bmm(q_vec.view(shape[0], 1, 3), v.view(
shape[0], 3, 1)).squeeze(-1) * 2.0
return a + b + c
@torch.jit.script
def quat_rotate_inverse(q, v):
shape = q.shape
q_w = q[:, -1]
q_vec = q[:, :3]
a = v * (2.0 * q_w ** 2 - 1.0).unsqueeze(-1)
b = torch.cross(q_vec, v, dim=-1) * q_w.unsqueeze(-1) * 2.0
c = q_vec * \
torch.bmm(q_vec.view(shape[0], 1, 3), v.view(
shape[0], 3, 1)).squeeze(-1) * 2.0
return a - b + c
@torch.jit.script
def quat_conjugate(a):
shape = a.shape
a = a.reshape(-1, 4)
return torch.cat((-a[:, :3], a[:, -1:]), dim=-1).view(shape)
@torch.jit.script
def quat_unit(a):
return normalize(a)
@torch.jit.script
def quat_from_angle_axis(angle, axis):
theta = (angle / 2).unsqueeze(-1)
xyz = normalize(axis) * theta.sin()
w = theta.cos()
return quat_unit(torch.cat([xyz, w], dim=-1))
@torch.jit.script
def normalize_angle(x):
return torch.atan2(torch.sin(x), torch.cos(x))
@torch.jit.script
def tf_inverse(q, t):
q_inv = quat_conjugate(q)
return q_inv, -quat_apply(q_inv, t)
@torch.jit.script
def tf_apply(q, t, v):
return quat_apply(q, v) + t
@torch.jit.script
def tf_vector(q, v):
return quat_apply(q, v)
@torch.jit.script
def tf_combine(q1, t1, q2, t2):
return quat_mul(q1, q2), quat_apply(q1, t2) + t1
@torch.jit.script
def get_basis_vector(q, v):
return quat_rotate(q, v)
def get_axis_params(value, axis_idx, x_value=0., dtype=np.float32, n_dims=3):
"""construct arguments to `Vec` according to axis index.
"""
zs = np.zeros((n_dims,))
assert axis_idx < n_dims, "the axis dim should be within the vector dimensions"
zs[axis_idx] = 1.
params = np.where(zs == 1., value, zs)
params[0] = x_value
return list(params.astype(dtype))
@torch.jit.script
def copysign(a, b):
# type: (float, Tensor) -> Tensor
a = torch.tensor(a, device=b.device, dtype=torch.float).repeat(b.shape[0])
return torch.abs(a) * torch.sign(b)
@torch.jit.script
def get_euler_xyz(q):
qx, qy, qz, qw = 0, 1, 2, 3
# roll (x-axis rotation)
sinr_cosp = 2.0 * (q[:, qw] * q[:, qx] + q[:, qy] * q[:, qz])
cosr_cosp = q[:, qw] * q[:, qw] - q[:, qx] * \
q[:, qx] - q[:, qy] * q[:, qy] + q[:, qz] * q[:, qz]
roll = torch.atan2(sinr_cosp, cosr_cosp)
# pitch (y-axis rotation)
sinp = 2.0 * (q[:, qw] * q[:, qy] - q[:, qz] * q[:, qx])
pitch = torch.where(
torch.abs(sinp) >= 1, copysign(np.pi / 2.0, sinp), torch.asin(sinp))
# yaw (z-axis rotation)
siny_cosp = 2.0 * (q[:, qw] * q[:, qz] + q[:, qx] * q[:, qy])
cosy_cosp = q[:, qw] * q[:, qw] + q[:, qx] * \
q[:, qx] - q[:, qy] * q[:, qy] - q[:, qz] * q[:, qz]
yaw = torch.atan2(siny_cosp, cosy_cosp)
return roll % (2 * np.pi), pitch % (2 * np.pi), yaw % (2 * np.pi)
@torch.jit.script
def quat_from_euler_xyz(roll, pitch, yaw):
cy = torch.cos(yaw * 0.5)
sy = torch.sin(yaw * 0.5)
cr = torch.cos(roll * 0.5)
sr = torch.sin(roll * 0.5)
cp = torch.cos(pitch * 0.5)
sp = torch.sin(pitch * 0.5)
qw = cy * cr * cp + sy * sr * sp
qx = cy * sr * cp - sy * cr * sp
qy = cy * cr * sp + sy * sr * cp
qz = sy * cr * cp - cy * sr * sp
return torch.stack([qx, qy, qz, qw], dim=-1)
@torch.jit.script
def torch_rand_float(lower, upper, shape, device):
# type: (float, float, Tuple[int, int], str) -> Tensor
return (upper - lower) * torch.rand(*shape, device=device) + lower
@torch.jit.script
def torch_random_dir_2(shape, device):
# type: (Tuple[int, int], str) -> Tensor
angle = torch_rand_float(-np.pi, np.pi, shape, device).squeeze(-1)
return torch.stack([torch.cos(angle), torch.sin(angle)], dim=-1)
@torch.jit.script
def tensor_clamp(t, min_t, max_t):
return torch.max(torch.min(t, max_t), min_t)
@torch.jit.script
def scale(x, lower, upper):
return (0.5 * (x + 1.0) * (upper - lower) + lower)
@torch.jit.script
def unscale(x, lower, upper):
return (2.0 * x - upper - lower) / (upper - lower)
def unscale_np(x, lower, upper):
return (2.0 * x - upper - lower) / (upper - lower)
def slope_platform_stairs_terrain(terrain,
slope=1,
step_width=0.2,
step_height=0.1,
num_steps=5):
"""
Generate a sloped followed by a platform and a downstair
Parameters:
terrain (SubTerrain): the terrain
slope (int): positive or negative slope
step_width (float): the width of the step [meters]
step_height (float): the height of the step [meters]
num_steps (int): number of the step of the stair
Returns:
terrain (SubTerrain): update terrain
"""
# switch parameters to discrete units
step_width = round(step_width / terrain.horizontal_scale)
step_height = round(step_height / terrain.vertical_scale)
height = step_height
for i in range(num_steps):
terrain.height_field_raw[i * step_width:(i + 1) *
step_width, :] += height
height += step_height
platform_height = height
slope_width = int(platform_height /
(np.abs(slope) *
(terrain.horizontal_scale / terrain.vertical_scale)))
x = np.arange(0, slope_width)
y = np.arange(0, terrain.length)
xx, yy = np.meshgrid(x, y, sparse=True)
xx = xx.reshape(slope_width, 1)
platform_width = int(terrain.width - num_steps * step_width - slope_width)
terrain.height_field_raw[num_steps * step_width:,
np.arange(terrain.length)] = platform_height
terrain.height_field_raw[-slope_width:,
np.arange(terrain.length)] -= (
platform_height * xx / slope_width).astype(
terrain.height_field_raw.dtype)
terrain.height_field_raw = terrain.height_field_raw[::-1]
return terrain
def stairs_platform_slope_terrain(terrain,
step_width=0.2,
step_height=0.1,
num_steps=10,
slope=0.26):
"""
Generate a stairs followed with a flat plane and a downslope
Parameters:
terrain (terrain): the terrain
step_width (float): the width of the step [meters]
step_height (float): the height of the step [meters]
num_steps (int): number of the step of the stair
slope (float): positive or negative slope
Returns:
terrain (SubTerrain): update terrain
"""
# switch parameters to discrete units
step_width = round(step_width / terrain.horizontal_scale)
step_height = round(step_height / terrain.vertical_scale)
height = step_height
for i in range(num_steps):
terrain.height_field_raw[i * step_width:(i + 1) *
step_width, :] += height
height += step_height
platform_height = height
slope_width = int(platform_height /
(np.abs(slope) *
(terrain.horizontal_scale / terrain.vertical_scale)))
x = np.arange(0, slope_width)
y = np.arange(0, terrain.length)
xx, yy = np.meshgrid(x, y, sparse=True)
xx = xx.reshape(slope_width, 1)
platform_width = int(terrain.width - num_steps * step_width - slope_width)
terrain.height_field_raw[num_steps * step_width:,
np.arange(terrain.length)] = platform_height
terrain.height_field_raw[-slope_width:,
np.arange(terrain.length)] -= (
platform_height * xx / slope_width).astype(
terrain.height_field_raw.dtype)
return terrain
if torch.cuda.is_available():
import string
import cupy as cp
def compute_meshes_normals_kernel():
compute_meshes_normals_kernel = cp.ElementwiseKernel(
in_params="raw U vertices, raw T triangles",
out_params="raw U output",
preamble=string.Template("""
__device__ bool cross(float p1x, float p1y, float p1z,
float p2x, float p2y, float p2z,
float& normalx,float& normaly,float& normalz)
{
float nx = p1y*p2z - p1z*p2y;
float ny = p1z*p2x - p1x*p2z;
float nz = p1x*p2y - p1y*p2x;
float l = sqrt(nx*nx + ny*ny + nz*nz);
if (l <1e-4) {
l = 1;
}
if (nz < 0.) {
nx *=-1;
ny *=-1;
nz *=-1;
}
atomicAdd(&normalx, nx/l);
atomicAdd(&normaly, ny/l);
atomicAdd(&normalz, nz/l);
return true;
}
""").substitute(),
operation=string.Template("""
int idx = i*2;
T mesh_1_0 = triangles[idx*3];
T mesh_1_1 = triangles[idx*3+1];
T mesh_1_2 = triangles[idx*3+2];
T mesh_2_0 = triangles[idx*3+3];
T mesh_2_1 = triangles[idx*3+4];
T mesh_2_2 = triangles[idx*3+5];
U p01x = vertices[mesh_1_1*3] - vertices[mesh_1_0*3];
U p01y = vertices[mesh_1_1*3+1] - vertices[mesh_1_0*3+1];
U p01z = vertices[mesh_1_1*3+2] - vertices[mesh_1_0*3+2];
U p02x = vertices[mesh_1_2*3] - vertices[mesh_1_0*3];
U p02y = vertices[mesh_1_2*3+1] - vertices[mesh_1_0*3+1];
U p02z = vertices[mesh_1_2*3+2] - vertices[mesh_1_0*3+2];
cross(p01x, p01y, p01z, p02x, p02y, p02z, output[3*i], output[3*i+1],output[3*i+2]);
p01x = vertices[mesh_2_1*3] - vertices[mesh_2_0*3];
p01y = vertices[mesh_2_1*3+1] - vertices[mesh_2_0*3+1];
p01z = vertices[mesh_2_1*3+2] - vertices[mesh_2_0*3+2];
p02x = vertices[mesh_2_2*3] - vertices[mesh_2_0*3];
p02y = vertices[mesh_2_2*3+1] - vertices[mesh_2_0*3+1];
p02z = vertices[mesh_2_2*3+2] - vertices[mesh_2_0*3+2];
cross(p01x, p01y, p01z, p02x, p02y, p02z, output[3*i], output[3*i+1],output[3*i+2]);
output[3*i] /= 2.0;
output[3*i+1] /= 2.0;
output[3*i+2] /= 2.0;
""").substitute(),
name="compute_meshes_normals_kernel",
)
return compute_meshes_normals_kernel
compute_meshes_normals_gpu_ = compute_meshes_normals_kernel()
def compute_meshes_normals_gpu(num_rows, num_cols, vertices, triangles):
vertices_gpu = cp.array(vertices, dtype=np.float32)
triangles_gpu = cp.array(triangles, dtype=np.int32)
output = cp.zeros((num_rows - 1, num_cols - 1, 3), dtype=np.float32)
compute_meshes_normals_gpu_(vertices_gpu,
triangles_gpu,
output,
size=((num_rows - 1) * (num_cols - 1)))
output_torch = torch.as_tensor(output)
return output_torch
def compute_meshes_normals(num_rows, num_cols, vertices, triangles):
"""
Compute normal vectors for all trimesh of the designed terrain
Parameters:
num_rows (int): Number of the row of the terrain
num_cols (int): Number of the column of the terrain
vertices (np.array(float)): array of shape (num_vertices, 3). Each row represents the location of each vertex [meters]
triangles (np.array(int)): array of shape (num_triangles, 3). Each row represents the indices of the 3 vertices connected by this triangle.
Returns:
normals (np.array(float)): array of shape (num_rows-1, num_cols-1, 3). Normals of the terrain
"""
normals = np.zeros((num_rows - 1, num_cols - 1, 3), dtype=np.float32)
for i in range(num_rows - 1):
start = 2 * i * (num_cols - 1)
for j in range(num_cols - 1):
idx_1 = start + 2 * j
mesh_1, mesh_2 = triangles[idx_1], triangles[idx_1 + 1]
normal1 = compute_triangle_normal(vertices[mesh_1[0]],
vertices[mesh_1[1]],
vertices[mesh_1[2]])
normal2 = compute_triangle_normal(vertices[mesh_2[0]],
vertices[mesh_2[1]],
vertices[mesh_2[2]])
new_normal = (
normal1 +
normal2) / 2.0 # np.mean(np.stack((normal1, normal2)), axis=0)
if new_normal[2] < 0.:
new_normal *= -1
normals[i][j] = new_normal
return torch.from_numpy(normals)
def compute_triangle_normal(p0, p1, p2):
"""
Compute the normal vector of a trimesh
Parameters:
p0 (np.array(float)): array of shape (3,). Position of the trimesh's first vertice
p1 (np.array(float)): array of shape (3,). Position of the trimesh's second vertice
p2 (np.array(float)): array of shape (3,). Position of the trimesh's third vertice
Returns:
normal (np.array(float)): array of shape (3,). Normal Vector of the given trimesh
"""
p0p1 = p1 - p0
p0p2 = p2 - p0
n = np.cross(p0p1, p0p2)
l = np.linalg.norm(n)
if l != 0.0:
normal = n / l
else:
normal = n
return normal
class Point:
def __init__(self, x, y, z):
self.x = x
self.y = y
self.z = z
@torch.jit.script
def get_euler_xyz(q):
qx, qy, qz, qw = 0, 1, 2, 3
# roll (x-axis rotation)
sinr_cosp = 2.0 * (q[:, qw] * q[:, qx] + q[:, qy] * q[:, qz])
cosr_cosp = q[:, qw] * q[:, qw] - q[:, qx] * \
q[:, qx] - q[:, qy] * q[:, qy] + q[:, qz] * q[:, qz]
roll = torch.atan2(sinr_cosp, cosr_cosp)
# pitch (y-axis rotation)
sinp = 2.0 * (q[:, qw] * q[:, qy] - q[:, qz] * q[:, qx])
pitch = torch.where(
torch.abs(sinp) >= 1, copysign(np.pi / 2.0, sinp), torch.asin(sinp))
# yaw (z-axis rotation)
siny_cosp = 2.0 * (q[:, qw] * q[:, qz] + q[:, qx] * q[:, qy])
cosy_cosp = q[:, qw] * q[:, qw] + q[:, qx] * \
q[:, qx] - q[:, qy] * q[:, qy] - q[:, qz] * q[:, qz]
yaw = torch.atan2(siny_cosp, cosy_cosp)
return torch.stack((roll, pitch, yaw), dim=-1)
@torch.jit.script
def coordinate_rotation(axis: int, angle):
s = torch.sin(angle)
c = torch.cos(angle)
R = torch.eye(3, dtype=torch.float, device=angle.device)
R = R.reshape((1, 3, 3)).repeat(angle.size(0), 1, 1)
if axis == 0:
R[:, 1, 1] = c
R[:, 2, 2] = c
R[:, 1, 2] = s
R[:, 2, 1] = -s
elif axis == 1:
R[:, 0, 0] = c
R[:, 0, 2] = -s
R[:, 2, 0] = s
R[:, 2, 2] = c
elif axis == 2:
R[:, 0, 0] = c
R[:, 0, 1] = s
R[:, 1, 0] = -s
R[:, 1, 1] = c
return R
@torch.jit.script
def euler_xyz_to_rotation_matrix(rpy):
N = rpy.size(0)
R = torch.tensor((N, 3, 3), dtype=torch.float, device=rpy.device)
R = torch.matmul(
torch.matmul(coordinate_rotation(0, rpy[:, 0]),
coordinate_rotation(1, rpy[:, 1])),
coordinate_rotation(2, rpy[:, 2]))
return R.transpose(1, 2)
def get_contact_normals(robots_feet_pos, mesh_normals, contact_forces):
pos_to_idx = robots_feet_pos[:, :, :2].to(torch.int)
pos_to_idx = torch.permute(pos_to_idx, (2, 0, 1))
pos_to_idx = torch.flatten(pos_to_idx, 1)
contact_normals = mesh_normals[pos_to_idx.tolist()]
contact_normals = contact_normals.view(-1, 4, 3)
filter = torch.norm(contact_forces[:, :, :], dim=2) > 1.
filter = torch.repeat_interleave(filter, repeats=3,
dim=1).view(-1, robots_feet_pos.shape[1],
3)
contact_normals *= filter
return contact_normals.flatten(1)
@torch.jit.script
def vector_2_skewmat(v):
N = v.size(0)
a = torch.zeros((N, 3, 3), device=v.device, dtype=torch.float)
a[:, 2, 1] = v[:, 0]
a[:, 2, 0] = -v[:, 1]
a[:, 1, 0] = v[:, 2]
a -= a.clone().transpose(1, 2)
return a
def parse_sim_params(args, cfg):
if 'real' in args.task:
return cfg
from isaacgym import gymapi, gymutil
# initialize sim params
sim_params = gymapi.SimParams()
if 'flex' in args.physics_engine:
args.physics_engine = gymapi.SIM_FLEX
if args.device != "cpu":
print("WARNING: Using Flex with GPU instead of PHYSX!")
else:
args.physics_engine = gymapi.SIM_PHYSX
sim_params.physx.use_gpu = True if 'cuda' in args.sim_device else False
sim_params.physx.num_subscenes = args.subscenes
sim_params.use_gpu_pipeline = True if 'cuda' in args.rl_device else False
# if sim options are provided in cfg, parse them and update/override above:
if "sim" in cfg:
gymutil.parse_sim_config(cfg["sim"], sim_params)
# Override num_threads if passed on the command line
if args.physics_engine == gymapi.SIM_PHYSX and args.num_threads > 0:
sim_params.physx.num_threads = args.num_threads
return sim_params