HiFAR/play_mujoco.py at main · Hi-FAR/HiFAR · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
import os
import sys
import glob
import yaml
import select
import argparse
import numpy as np
import time
import torch
import mujoco, mujoco.viewer
from utils.model import *


class MujocoRunner:

    def __init__(self):
        self.get_args()
        self.update_cfg_from_args()
        self.model = DenoisingRMA(
            self.cfg["env"]["num_actions"], self.cfg["env"]["num_observations"], self.cfg["runner"]["num_stack"], self.cfg["env"]["num_privileged_obs"], self.cfg["algorithm"]["num_embedding"]
        )
        self.load()
        self.mj_model = mujoco.MjModel.from_xml_path("resources/T1/T1_serial_collision.xml")
        self.mj_model.opt.timestep = self.cfg["sim"]["dt"] / self.cfg["sim"]["substeps"]
        self.mj_data = mujoco.MjData(self.mj_model)
        mujoco.mj_resetData(self.mj_model, self.mj_data)
        self.mj_data.qpos = np.array([0, 0, 0.25, 0.7071, 0, -0.7071, 0, 0.2, -1.35, 0, -0.5, 0.2, 1.35, 0, 0.5, 0, -0.2, 0, 0, 0.4, -0.25, 0, -0.2, 0, 0, 0.4, -0.25, 0], dtype=np.float32)
        mujoco.mj_forward(self.mj_model, self.mj_data)

    def get_args(self):
        parser = argparse.ArgumentParser()
        parser.add_argument("--task", required=True, type=str, help="Name of the task to run.")
        parser.add_argument("--checkpoint", type=str, help="Path of model checkpoint to load. Overrides config file if provided.")
        self.args = parser.parse_args()

    def update_cfg_from_args(self):
        cfg_file = os.path.join("envs", "{}.yaml".format(self.args.task))
        with open(cfg_file, "r", encoding="utf-8") as f:
            self.cfg = yaml.load(f.read(), Loader=yaml.FullLoader)
        for arg in vars(self.args):
            if getattr(self.args, arg) is not None:
                self.cfg["basic"][arg] = getattr(self.args, arg)

    def load(self):
        if not self.cfg["basic"]["checkpoint"]:
            raise ValueError("Invalid model checkpoint")
        if (self.cfg["basic"]["checkpoint"] == "-1") or (self.cfg["basic"]["checkpoint"] == -1):
            self.cfg["basic"]["checkpoint"] = sorted(glob.glob(os.path.join("logs", "**/*.pth"), recursive=True), key=os.path.getmtime)[-1]
        print("Loading model from {}".format(self.cfg["basic"]["checkpoint"]))
        model_dict = torch.load(self.cfg["basic"]["checkpoint"], map_location="cpu", weights_only=True)
        self.model.load_state_dict(model_dict["model"])

    def play(self):
        default_dof_pos = np.array([0.2, -1.35, 0, -0.5, 0.2, 1.35, 0, 0.5, 0, -0.2, 0, 0, 0.4, -0.25, 0, -0.2, 0, 0, 0.4, -0.25, 0], dtype=np.float32)
        stiffness = np.array([200, 200, 200, 200, 200, 200, 200, 200, 200, 200, 200, 200, 200, 100, 100, 200, 200, 200, 200, 100, 100], dtype=np.float32)
        damping = np.array([2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 2, 2, 2, 2, 1, 1], dtype=np.float32)
        # ctrl_limit = np.array([30, 30, 30, 30, 30, 30, 30, 30, 30, 45, 45, 30, 65, 24, 6, 45, 45, 30, 65, 24, 6], dtype=np.float32)
        ctrl_limit = np.array([24, 24, 24, 24, 24, 24, 24, 24, 30, 45, 45, 30, 65, 24, 6, 45, 45, 30, 65, 24, 6], dtype=np.float32)
        # ctrl_limit = np.array([18, 18, 18, 18, 18, 18, 18, 18, 30, 45, 45, 30, 65, 24, 6, 45, 45, 30, 65, 24, 6], dtype=np.float32)
        actions = np.zeros((self.cfg["env"]["num_actions"]), dtype=np.float32)
        dof_targets = np.zeros(default_dof_pos.shape, dtype=np.float32)
        stacked_obs = np.zeros((1, self.cfg["runner"]["num_stack"], self.cfg["env"]["num_observations"]), dtype=np.float32)
        effective_ids = [0, 3, 4, 7, 9, 12, 13, 15, 18, 19]

        it = 0
        with mujoco.viewer.launch_passive(self.mj_model, self.mj_data) as viewer:
            viewer.cam.elevation = -20
            while viewer.is_running():
                if it % 2500 == 2499:
                    body_id = mujoco.mj_name2id(self.mj_model, mujoco.mjtObj.mjOBJ_BODY, "Trunk")
                    external_force = 30000 * np.random.rand(2)
                    self.mj_data.xfrc_applied[body_id, :2] = external_force
                    print(f"Applied external force: {external_force} at step {it}")

                dof_pos = self.mj_data.qpos.astype(np.float32)[7:]
                dof_vel = self.mj_data.qvel.astype(np.float32)[6:]
                quat = self.mj_data.sensor("orientation").data[[1, 2, 3, 0]].astype(np.float32)
                base_ang_vel = self.mj_data.sensor("angular-velocity").data.astype(np.float32)
                projected_gravity = self.quat_rotate_inverse(quat, np.array([0.0, 0.0, -1.0]))
                if it % (self.cfg["control"]["decimation"] * self.cfg["sim"]["substeps"]) == 0:
                    obs = np.zeros([1, self.cfg["env"]["num_observations"]], dtype=np.float32)
                    obs[0, 0:3] = projected_gravity * self.cfg["normalization"]["gravity"]
                    obs[0, 3:6] = base_ang_vel * self.cfg["normalization"]["ang_vel"]
                    obs[0, 6:16] = (dof_pos - default_dof_pos)[effective_ids] * self.cfg["normalization"]["dof_pos"]
                    obs[0, 16:26] = dof_vel[effective_ids] * self.cfg["normalization"]["dof_vel"]
                    obs[0, 26:36] = actions
                    if it == 0:
                        stacked_obs[:] = obs[:, np.newaxis, :]
                    stacked_obs[:, 1:, :] = stacked_obs[:, :-1, :]
                    stacked_obs[:, 0, :] = obs
                    mirrored_obs = self.mirror_obs(obs)
                    mirrored_stacked_obs = self.mirror_obs(stacked_obs)
                    dist, _ = self.model.act(torch.tensor(obs), stacked_obs=torch.tensor(stacked_obs))
                    mirrored_dist, _ = self.model.act(torch.tensor(mirrored_obs), stacked_obs=torch.tensor(mirrored_stacked_obs))
                    actions[:] = 0.5 * (dist.loc.detach().numpy() + self.mirror_act(mirrored_dist.loc.detach().numpy()))
                    actions = np.clip(actions, -self.cfg["normalization"]["clip_actions"], self.cfg["normalization"]["clip_actions"])
                    dof_targets[:] = default_dof_pos
                    dof_targets[effective_ids] += self.cfg["control"]["action_scale"] * actions
                self.mj_data.ctrl = np.clip(stiffness * (dof_targets - dof_pos) - damping * dof_vel, -ctrl_limit, ctrl_limit)
                mujoco.mj_step(self.mj_model, self.mj_data)
                viewer.cam.lookat[:] = self.mj_data.qpos.astype(np.float32)[0:3]
                viewer.sync()
                time.sleep(self.cfg["sim"]["dt"] / self.cfg["sim"]["substeps"])
                it += 1

    @staticmethod
    def quat_rotate_inverse(q, v):
        q_w = q[-1]
        q_vec = q[:3]
        a = v * (2.0 * q_w**2 - 1.0)
        b = np.cross(q_vec, v) * (q_w * 2.0)
        c = q_vec * (np.dot(q_vec, v) * 2.0)
        return a - b + c

    @staticmethod
    def mirror_obs(obs):
        mat = np.zeros((36, 36), dtype=np.float32)
        mat[ 0: 6,  0: 6] = np.eye(6)
        mat[ 6: 8,  8:10] = np.eye(2)
        mat[ 8:10,  6: 8] = np.eye(2)
        mat[10:13, 13:16] = np.eye(3)
        mat[13:16, 10:13] = np.eye(3)
        mat[16:18, 18:20] = np.eye(2)
        mat[18:20, 16:18] = np.eye(2)
        mat[20:23, 23:26] = np.eye(3)
        mat[23:26, 20:23] = np.eye(3)
        mat[26:28, 28:30] = np.eye(2)
        mat[28:30, 26:28] = np.eye(2)
        mat[30:33, 33:36] = np.eye(3)
        mat[33:36, 30:33] = np.eye(3)
        flip_val = np.ones(36, dtype=np.float32)
        inverse_ids = [  1,  3,  5,
                         7,  9,
                        17, 19,
                        27, 29]
        flip_val[inverse_ids] = -1
        flip_mat = np.diag(flip_val)
        mirror_transform_mat = np.dot(mat, flip_mat)

        orig_shape = obs.shape
        reshaped_obs = obs.reshape(-1, 36)
        mirrored_obs = np.dot(reshaped_obs, mirror_transform_mat.T)
        mirrored_obs = mirrored_obs.reshape(orig_shape)
        return mirrored_obs

    @staticmethod
    def mirror_act(act):
        mat = np.zeros((10, 10), dtype=np.float32)
        mat[0:2, 2:4] = np.eye(2)
        mat[2:4, 0:2] = np.eye(2)
        mat[4:7, 7:10] = np.eye(3)
        mat[7:10, 4:7] = np.eye(3)
        flip_val = np.ones(10, dtype=np.float32)
        inverse_ids = [1, 3]
        flip_val[inverse_ids] = -1
        flip_mat = np.diag(flip_val)
        mirror_transform_mat = np.dot(mat, flip_mat)
        mirrored_act = np.dot(act, mirror_transform_mat.T)
        return mirrored_act


if __name__ == "__main__":
    runner = MujocoRunner()
    runner.play()