# Copyright (c) 2021-2023, NVIDIA Corporation
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
#
# 3. Neither the name of the copyright holder nor the names of its
# contributors may be used to endorse or promote products derived from
# this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE..
import os
from enum import Enum
from gym import spaces
from isaacgym import gymtorch
import rofunc as rf
from rofunc.learning.RofuncRL.tasks.isaacgymenv.amp.humanoid_amp_base import HumanoidAMPBase, dof_to_obs
from rofunc.learning.RofuncRL.tasks.isaacgymenv.amp.motion_lib import MotionLib
from rofunc.learning.RofuncRL.tasks.utils.torch_jit_utils import *
NUM_AMP_OBS_PER_STEP = 13 + 52 + 28 + 12 # [root_h, root_rot, root_vel, root_ang_vel, dof_pos, dof_vel, key_body_pos]
[docs]class HumanoidAMPTask(HumanoidAMPBase):
[docs] class StateInit(Enum):
Default = 0
Start = 1
Random = 2
Hybrid = 3
def __init__(self, cfg, rl_device, sim_device, graphics_device_id, headless, virtual_screen_capture, force_render):
self.cfg = cfg
state_init = cfg["env"]["stateInit"]
self._state_init = HumanoidAMPTask.StateInit[state_init]
self._hybrid_init_prob = cfg["env"]["hybridInitProb"]
self._num_amp_obs_steps = cfg["env"]["numAMPObsSteps"]
assert (self._num_amp_obs_steps >= 2)
self._reset_default_env_ids = []
self._reset_ref_env_ids = []
super().__init__(config=self.cfg, rl_device=rl_device, sim_device=sim_device,
graphics_device_id=graphics_device_id, headless=headless,
virtual_screen_capture=virtual_screen_capture, force_render=force_render)
motion_file = cfg['env'].get('motion_file', "amp_humanoid_backflip.npy")
if rf.oslab.is_absl_path(motion_file):
motion_file_path = motion_file
else:
motion_file_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
"../../../../../examples/data/amp/" + motion_file)
self._load_motion(motion_file_path)
self.num_amp_obs = self._num_amp_obs_steps * NUM_AMP_OBS_PER_STEP
self._amp_obs_space = spaces.Box(np.ones(self.num_amp_obs) * -np.Inf, np.ones(self.num_amp_obs) * np.Inf)
self._amp_obs_buf = torch.zeros((self.num_envs, self._num_amp_obs_steps, NUM_AMP_OBS_PER_STEP),
device=self.device, dtype=torch.float)
self._curr_amp_obs_buf = self._amp_obs_buf[:, 0]
self._hist_amp_obs_buf = self._amp_obs_buf[:, 1:]
self._amp_obs_demo_buf = None
return
[docs] def post_physics_step(self):
super().post_physics_step()
self._update_hist_amp_obs()
self._compute_amp_observations()
amp_obs_flat = self._amp_obs_buf.view(-1, self.get_num_amp_obs())
self.extras["amp_obs"] = amp_obs_flat
return
[docs] def get_num_amp_obs(self):
return self.num_amp_obs
@property
def amp_observation_space(self):
return self._amp_obs_space
# def fetch_amp_obs_demo(self, num_samples):
# return self.task.fetch_amp_obs_demo(num_samples)
[docs] def fetch_amp_obs_demo(self, num_samples):
dt = self.dt
motion_ids = self._motion_lib.sample_motions(num_samples) # [1024] all 0
if self._amp_obs_demo_buf is None: # [1024, 2, 105]
self._build_amp_obs_demo_buf(num_samples)
else:
assert (self._amp_obs_demo_buf.shape[0] == num_samples)
motion_times0 = self._motion_lib.sample_time(motion_ids) # [1024] all float time
motion_ids = np.tile(np.expand_dims(motion_ids, axis=-1), [1, self._num_amp_obs_steps]) # [1024, 2]
motion_times = np.expand_dims(motion_times0, axis=-1) # [1024, 1]
time_steps = -dt * np.arange(0, self._num_amp_obs_steps) # [_num_amp_obs_steps] _num_amp_obs_steps=2 [-0. -0.0332]
motion_times = motion_times + time_steps # [1024, 2] [14.17973511 14.14653511 7.12111682 7.08791682 ...
motion_ids = motion_ids.flatten() # [2048]
motion_times = motion_times.flatten() # [2048]
root_pos, root_rot, dof_pos, root_vel, root_ang_vel, dof_vel, key_pos \
= self._motion_lib.get_motion_state(motion_ids, motion_times)
root_states = torch.cat([root_pos, root_rot, root_vel, root_ang_vel], dim=-1)
amp_obs_demo = build_amp_observations(root_states, dof_pos, dof_vel, key_pos,
self._local_root_obs) # [2048, 105]
self._amp_obs_demo_buf[:] = amp_obs_demo.view(self._amp_obs_demo_buf.shape) # [1024, 2, 105]
amp_obs_demo_flat = self._amp_obs_demo_buf.view(-1, self.get_num_amp_obs()) # [1024, 210]
return amp_obs_demo_flat
def _build_amp_obs_demo_buf(self, num_samples):
self._amp_obs_demo_buf = torch.zeros((num_samples, self._num_amp_obs_steps, NUM_AMP_OBS_PER_STEP),
device=self.device, dtype=torch.float)
return
def _load_motion(self, motion_file):
self._motion_lib = MotionLib(motion_file=motion_file,
num_dofs=self.num_dof,
key_body_ids=self._key_body_ids.cpu().numpy(),
device=self.device)
return
[docs] def reset_idx(self, env_ids):
super().reset_idx(env_ids)
self._init_amp_obs(env_ids)
return
def _reset_actors(self, env_ids):
if self._state_init == HumanoidAMPTask.StateInit.Default:
self._reset_default(env_ids)
elif self._state_init == HumanoidAMPTask.StateInit.Start \
or self._state_init == HumanoidAMPTask.StateInit.Random:
self._reset_ref_state_init(env_ids)
elif self._state_init == HumanoidAMPTask.StateInit.Hybrid:
self._reset_hybrid_state_init(env_ids)
else:
assert False, "Unsupported state initialization strategy: {:s}".format(str(self._state_init))
self.progress_buf[env_ids] = 0
self.reset_buf[env_ids] = 0
self._terminate_buf[env_ids] = 0
return
def _reset_default(self, env_ids):
self._dof_pos[env_ids] = self._initial_dof_pos[env_ids]
self._dof_vel[env_ids] = self._initial_dof_vel[env_ids]
env_ids_int32 = env_ids.to(dtype=torch.int32)
self.gym.set_actor_root_state_tensor_indexed(self.sim, gymtorch.unwrap_tensor(self._initial_root_states),
gymtorch.unwrap_tensor(env_ids_int32), len(env_ids_int32))
self.gym.set_dof_state_tensor_indexed(self.sim, gymtorch.unwrap_tensor(self._dof_state),
gymtorch.unwrap_tensor(env_ids_int32), len(env_ids_int32))
self._reset_default_env_ids = env_ids
return
def _reset_ref_state_init(self, env_ids):
num_envs = env_ids.shape[0]
motion_ids = self._motion_lib.sample_motions(num_envs)
if self._state_init == HumanoidAMPTask.StateInit.Random \
or self._state_init == HumanoidAMPTask.StateInit.Hybrid:
motion_times = self._motion_lib.sample_time(motion_ids)
elif self._state_init == HumanoidAMPTask.StateInit.Start:
motion_times = np.zeros(num_envs)
else:
assert False, "Unsupported state initialization strategy: {:s}".format(str(self._state_init))
root_pos, root_rot, dof_pos, root_vel, root_ang_vel, dof_vel, key_pos \
= self._motion_lib.get_motion_state(motion_ids, motion_times)
self._set_env_state(env_ids=env_ids,
root_pos=root_pos,
root_rot=root_rot,
dof_pos=dof_pos,
root_vel=root_vel,
root_ang_vel=root_ang_vel,
dof_vel=dof_vel)
self._reset_ref_env_ids = env_ids
self._reset_ref_motion_ids = motion_ids
self._reset_ref_motion_times = motion_times
return
def _reset_hybrid_state_init(self, env_ids):
num_envs = env_ids.shape[0]
ref_probs = to_torch(np.array([self._hybrid_init_prob] * num_envs), device=self.device)
ref_init_mask = torch.bernoulli(ref_probs) == 1.0
ref_reset_ids = env_ids[ref_init_mask]
if len(ref_reset_ids) > 0:
self._reset_ref_state_init(ref_reset_ids)
default_reset_ids = env_ids[torch.logical_not(ref_init_mask)]
if len(default_reset_ids) > 0:
self._reset_default(default_reset_ids)
return
def _init_amp_obs(self, env_ids):
self._compute_amp_observations(env_ids)
if len(self._reset_default_env_ids) > 0:
self._init_amp_obs_default(self._reset_default_env_ids)
if len(self._reset_ref_env_ids) > 0:
self._init_amp_obs_ref(self._reset_ref_env_ids, self._reset_ref_motion_ids,
self._reset_ref_motion_times)
return
def _init_amp_obs_default(self, env_ids):
curr_amp_obs = self._curr_amp_obs_buf[env_ids].unsqueeze(-2)
self._hist_amp_obs_buf[env_ids] = curr_amp_obs
return
def _init_amp_obs_ref(self, env_ids, motion_ids, motion_times):
dt = self.dt
motion_ids = np.tile(np.expand_dims(motion_ids, axis=-1), [1, self._num_amp_obs_steps - 1])
motion_times = np.expand_dims(motion_times, axis=-1)
time_steps = -dt * (np.arange(0, self._num_amp_obs_steps - 1) + 1)
motion_times = motion_times + time_steps
motion_ids = motion_ids.flatten()
motion_times = motion_times.flatten()
root_pos, root_rot, dof_pos, root_vel, root_ang_vel, dof_vel, key_pos \
= self._motion_lib.get_motion_state(motion_ids, motion_times)
root_states = torch.cat([root_pos, root_rot, root_vel, root_ang_vel], dim=-1)
amp_obs_demo = build_amp_observations(root_states, dof_pos, dof_vel, key_pos,
self._local_root_obs)
self._hist_amp_obs_buf[env_ids] = amp_obs_demo.view(self._hist_amp_obs_buf[env_ids].shape)
return
def _set_env_state(self, env_ids, root_pos, root_rot, dof_pos, root_vel, root_ang_vel, dof_vel):
self._root_states[env_ids, 0:3] = root_pos
self._root_states[env_ids, 3:7] = root_rot
self._root_states[env_ids, 7:10] = root_vel
self._root_states[env_ids, 10:13] = root_ang_vel
self._dof_pos[env_ids] = dof_pos
self._dof_vel[env_ids] = dof_vel
env_ids_int32 = env_ids.to(dtype=torch.int32)
self.gym.set_actor_root_state_tensor_indexed(self.sim, gymtorch.unwrap_tensor(self._root_states),
gymtorch.unwrap_tensor(env_ids_int32), len(env_ids_int32))
self.gym.set_dof_state_tensor_indexed(self.sim, gymtorch.unwrap_tensor(self._dof_state),
gymtorch.unwrap_tensor(env_ids_int32), len(env_ids_int32))
return
def _update_hist_amp_obs(self, env_ids=None):
if env_ids is None:
for i in reversed(range(self._amp_obs_buf.shape[1] - 1)):
self._amp_obs_buf[:, i + 1] = self._amp_obs_buf[:, i]
else:
for i in reversed(range(self._amp_obs_buf.shape[1] - 1)):
self._amp_obs_buf[env_ids, i + 1] = self._amp_obs_buf[env_ids, i]
return
def _compute_amp_observations(self, env_ids=None):
key_body_pos = self._rigid_body_pos[:, self._key_body_ids, :]
if env_ids is None:
self._curr_amp_obs_buf[:] = build_amp_observations(self._root_states, self._dof_pos, self._dof_vel,
key_body_pos, self._local_root_obs)
else:
self._curr_amp_obs_buf[env_ids] = build_amp_observations(self._root_states[env_ids], self._dof_pos[env_ids],
self._dof_vel[env_ids], key_body_pos[env_ids],
self._local_root_obs)
return
#####################################################################
###=========================jit functions=========================###
#####################################################################
@torch.jit.script
def build_amp_observations(root_states, dof_pos, dof_vel, key_body_pos, local_root_obs):
# type: (Tensor, Tensor, Tensor, Tensor, bool) -> Tensor
root_pos = root_states[:, 0:3]
root_rot = root_states[:, 3:7]
root_vel = root_states[:, 7:10]
root_ang_vel = root_states[:, 10:13]
root_h = root_pos[:, 2:3]
heading_rot = calc_heading_quat_inv(root_rot)
if local_root_obs:
root_rot_obs = quat_mul(heading_rot, root_rot)
else:
root_rot_obs = root_rot
root_rot_obs = quat_to_tan_norm(root_rot_obs)
local_root_vel = my_quat_rotate(heading_rot, root_vel)
local_root_ang_vel = my_quat_rotate(heading_rot, root_ang_vel)
root_pos_expand = root_pos.unsqueeze(-2)
local_key_body_pos = key_body_pos - root_pos_expand
heading_rot_expand = heading_rot.unsqueeze(-2)
heading_rot_expand = heading_rot_expand.repeat((1, local_key_body_pos.shape[1], 1))
flat_end_pos = local_key_body_pos.view(local_key_body_pos.shape[0] * local_key_body_pos.shape[1],
local_key_body_pos.shape[2])
flat_heading_rot = heading_rot_expand.view(heading_rot_expand.shape[0] * heading_rot_expand.shape[1],
heading_rot_expand.shape[2])
local_end_pos = my_quat_rotate(flat_heading_rot, flat_end_pos)
flat_local_key_pos = local_end_pos.view(local_key_body_pos.shape[0],
local_key_body_pos.shape[1] * local_key_body_pos.shape[2])
dof_obs = dof_to_obs(dof_pos)
obs = torch.cat((root_h, root_rot_obs, local_root_vel, local_root_ang_vel, dof_obs, dof_vel, flat_local_key_pos),
dim=-1)
return obs
