Hugging Face's logo

Raghav-16
/
Microduck_Manipulation

Robotics
ONNX
mujoco
reinforcement-learning
biomimetic-control
bipedal-locomotion
simulation
Model card Files
xet
 Community
Microduck_Manipulation
File size: 6,844 Bytes
dad190e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import sys, os, math
from pathlib import Path
import numpy as np
import mujoco
import imageio

ws = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(ws / "scripts"))
from test_spoon_lemon_balance import create_spoon_lemon_scene_xml, create_policy_with_steadicam

sim_duration_sec = 25.0  # 25.0 seconds at authentic 1x natural physics time (>= 20s)
target_vx = 0.20
fps = 25

mp4_path = str(ws / "docs" / "media" / "lemon_spoon_balance.mp4")
webp_path = str(ws / "docs" / "media" / "lemon_spoon_balance.webp")
policy_path = str(ws / "policies" / "onnx" / "payload_clamping_locomotion.onnx")

print(f"[1X LEMON] Initializing simulation with proven smooth policy {policy_path}...")
xml_str = create_spoon_lemon_scene_xml(0.50)
model = mujoco.MjModel.from_xml_string(xml_str)
model.opt.timestep = 0.002
data = mujoco.MjData(model)

policy, is_rl = create_policy_with_steadicam(model, data, policy_path)
policy.head_pos_steadicam = 0.0

spoon_bid = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_BODY, "spoon")
lemon_bid = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_BODY, "lemon")
beak_jid = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_JOINT, "beak")
beak_dof = model.jnt_dofadr[beak_jid]
beak_qpos = model.jnt_qposadr[beak_jid]
jaw_bid = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_BODY, "jaw_soft")
spoon_jid = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_JOINT, "spoon_free")
spoon_qadr = model.jnt_qposadr[spoon_jid]
lemon_jid = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_JOINT, "lemon_free")
lemon_qadr = model.jnt_qposadr[lemon_jid]

mujoco.mj_resetData(model, data)
data.qpos[0:3] = [0.0, 0.0, 0.125]
data.qpos[3:7] = [1.0, 0.0, 0.0, 0.0]
for i, idx in enumerate(policy.joint_qpos_indices):
    data.qpos[idx] = policy.default_pose[i]
data.qpos[beak_qpos] = 1.550
data.qfrc_applied[beak_dof] = -0.45
mujoco.mj_forward(model, data)

p_jaw = data.xpos[jaw_bid]
R_jaw = data.xmat[jaw_bid].reshape(3, 3)
data.qpos[spoon_qadr:spoon_qadr+3] = p_jaw + R_jaw @ np.array([-0.0119, 0.0, -0.0465])
q_jaw = np.empty(4); mujoco.mju_mat2Quat(q_jaw, data.xmat[jaw_bid])
q_s = np.empty(4); mujoco.mju_mulQuat(q_s, q_jaw, np.array([0.0, 0.7071068, 0.7071068, 0.0]))
data.qpos[spoon_qadr+3:spoon_qadr+7] = q_s
mujoco.mj_forward(model, data)

p_spoon = data.xpos[spoon_bid]
R_spoon = data.xmat[spoon_bid].reshape(3, 3)
data.qpos[lemon_qadr:lemon_qadr+3] = p_spoon + R_spoon @ np.array([0.0, 0.0052, 0.070])
data.qpos[lemon_qadr+3:lemon_qadr+7] = q_s
mujoco.mj_forward(model, data)

# Settle 150 steps
for _ in range(150):
    data.qpos[0:3] = [0.0, 0.0, 0.125]
    data.qvel[0:6] = 0.0
    data.qfrc_applied[beak_dof] = float(np.clip(-0.45 + 10.0 * (1.550 - data.qpos[beak_qpos]) - 0.3 * data.qvel[beak_dof], -1.0, 0.2))
    mujoco.mj_step(model, data)

policy.set_vel_cmd(target_vx, 0.0, 0.0)

# Pre-step 200 physics steps so robot is actively walking in frame 0 (NO static pause)
for step in range(200):
    data.qfrc_applied[beak_dof] = float(np.clip(-0.45 + 10.0 * (1.550 - data.qpos[beak_qpos]) - 0.3 * data.qvel[beak_dof], -1.0, 0.2))
    if step % 10 == 0:
        act = policy.infer()
        R_trunk = data.xmat[1].reshape(3, 3)
        h_fwd = np.array([R_trunk[0, 0], R_trunk[1, 0], 0.0])
        h_fwd /= (np.linalg.norm(h_fwd) + 1e-6)
        R_s = data.xmat[spoon_bid].reshape(3, 3)
        b_norm = R_s @ np.array([0.0, 1.0, 0.0])
        spoon_pitch = math.asin(np.clip(float(np.dot(b_norm, h_fwd)), -0.999, 0.999))
        rate_cmd = np.clip(1.20 * spoon_pitch, -0.30, 0.30)
        policy.head_pos_steadicam = np.clip(policy.head_pos_steadicam + rate_cmd * 0.02, -0.40, 0.40)
        act[5] = 0.0 + policy.head_pos_steadicam
        act[6] = 0.0 - policy.head_pos_steadicam
        act[7] = 0.0
        act[8] = 0.0
        policy.apply_action(act)
    mujoco.mj_step(model, data)

renderer = mujoco.Renderer(model, width=1280, height=720)
vopt = mujoco.MjvOption()
vopt.geomgroup[3] = 0
cam = mujoco.MjvCamera()
cam.type = mujoco.mjtCamera.mjCAMERA_TRACKING
cam.trackbodyid = 1
cam.distance = 0.72
cam.elevation = -14.0
cam.azimuth = 138.0

total_sim_steps = int(sim_duration_sec / model.opt.timestep)
sim_steps_per_frame = int(round(1.0 / (fps * model.opt.timestep)))

# Write directly to mp4_path at 1x real-time natural physics speed (25 fps for 25.0s = 625 frames)
writer = imageio.get_writer(mp4_path, fps=fps, codec="libx264", quality=8)
frame_count = 0

print(f"[1X LEMON] Recording {sim_duration_sec:.1f}s at authentic 1X natural speed ({total_sim_steps} steps, {fps} fps)...", flush=True)
for step in range(total_sim_steps):
    t_sim = step * model.opt.timestep
    data.qfrc_applied[beak_dof] = float(np.clip(-0.45 + 10.0 * (1.550 - data.qpos[beak_qpos]) - 0.3 * data.qvel[beak_dof], -1.0, 0.2))

    if step % 10 == 0:
        act = policy.infer()
        R_trunk = data.xmat[1].reshape(3, 3)
        h_fwd = np.array([R_trunk[0, 0], R_trunk[1, 0], 0.0])
        h_fwd /= (np.linalg.norm(h_fwd) + 1e-6)
        R_s = data.xmat[spoon_bid].reshape(3, 3)
        b_norm = R_s @ np.array([0.0, 1.0, 0.0])
        spoon_pitch = math.asin(np.clip(float(np.dot(b_norm, h_fwd)), -0.999, 0.999))
        rate_cmd = np.clip(1.20 * spoon_pitch, -0.30, 0.30)
        policy.head_pos_steadicam = np.clip(policy.head_pos_steadicam + rate_cmd * 0.02, -0.40, 0.40)
        act[5] = 0.0 + policy.head_pos_steadicam
        act[6] = 0.0 - policy.head_pos_steadicam
        act[7] = 0.0
        act[8] = 0.0
        policy.apply_action(act)

    mujoco.mj_step(model, data)

    if step % sim_steps_per_frame == 0:
        renderer.update_scene(data, camera=cam, scene_option=vopt)
        writer.append_data(renderer.render())
        frame_count += 1
        if frame_count % (fps * 5) == 0:
            p_s = data.xpos[spoon_bid]
            R_s = data.xmat[spoon_bid].reshape(3, 3)
            b_c = p_s + R_s @ np.array([0.0, 0.0052, 0.070])
            dev = np.linalg.norm(data.xpos[lemon_bid] - b_c) * 1000.0
            print(f"  Frame {frame_count:4d}/{int(sim_duration_sec*fps)} (t={t_sim:5.1f}s): walked {data.qpos[0]:+5.2f}m | lemon dev={dev:4.1f}mm", flush=True)

writer.close()
print(f"[1X LEMON SUCCESS] Saved 1x MP4 ({mp4_path}) - {frame_count} frames, duration = {frame_count/fps:.2f}s", flush=True)

# Generate high-resolution 600px wide WebP animation at 1x natural playback speed
cmd_webp = f'ffmpeg -y -i "{mp4_path}" -vf "fps=10,scale=600:-1:flags=lanczos" -vcodec libwebp -lossless 0 -compression_level 4 -q:v 38 -loop 0 "{webp_path}"'
os.system(cmd_webp)

root_mp4 = str(ws / "docs" / "media" / "lemon_spoon_balance.mp4")
root_webp = str(ws / "docs" / "media" / "lemon_spoon_balance.webp")
os.system(f'cp "{mp4_path}" "{root_mp4}"')
os.system(f'cp "{webp_path}" "{root_webp}"')

sz = os.path.getsize(webp_path) / 1e6
print(f"[1X LEMON SUCCESS] Saved 1x WebP ({webp_path}) - Size: {sz:.2f} MB", flush=True)