catachiii's picture
Initial release: 500 tennis strokes for the Unitree G1
390e994 verified
|
Raw History Blame Contribute Delete
4.24 kB

FiveHundredShots visualizer

visualize.py replays the clips on the G1 + racket model in the MuJoCo viewer, or renders them to a video file without a display. fivehundredshots.py is the loader it uses; it needs only numpy and is the place to look for the clip layout.

Install

pip install -r scripts/requirements.txt    # numpy, mujoco, imageio[ffmpeg]

Python 3.9 or newer. imageio[ffmpeg] is only needed for --record.

Quick start

Run from the dataset folder:

python scripts/visualize.py                            # browse all 500 shots
python scripts/visualize.py --stroke backhand          # only backhands
python scripts/visualize.py --shot shot_0458 shot_0122 # only these shots
python scripts/visualize.py --shot shot_0458 --speed 0.25   # quarter speed

On macOS the interactive viewer has to be started with mjpython instead of python (a MuJoCo requirement).

What you see

forehand

Marker Meaning
Yellow sphere Centre of the racket string bed in the current frame
Red sphere Where that centre is at ball contact (sweetspot_hit_world)
Green arrow Direction the racket face points at contact (hit_direction_world)

The yellow sphere reaches the red one exactly at the contact frame. The camera follows the pelvis and is placed relative to each shot's hit direction, so every stroke is seen from the same side; the mouse still orbits and zooms. The top-left label shows the shot id, stroke, current frame and contact frame.

Keyboard

Key Action
SPACE Pause / resume
RIGHT or N Next shot
LEFT or B Previous shot
R Restart the shot
H Jump to the contact frame and pause
V Show / hide the markers
UP / DOWN Faster / slower playback

Each shot loops until you move on. Close the window to quit.

Record a video

python scripts/visualize.py --shot shot_0458 shot_0122 --record strokes.mp4
python scripts/visualize.py --shot shot_0458 --record slow.mp4 --speed 0.25
python scripts/visualize.py --shot shot_0458 --record small.gif --every 2 --width 480 --height 360

--record needs --shot and renders the listed shots one after another, with the same camera and markers as the viewer. Every rendered frame is a frame of the dataset: --speed only changes the frame rate written to the file, and --every N keeps every N-th frame.

On a machine without a display, select an offscreen backend first:

MUJOCO_GL=egl python scripts/visualize.py --shot shot_0458 --record out.mp4

Options

Option Default Meaning
--shot ID [ID ...] all Play only these shots
--stroke forehand|backhand both Play only this stroke type
--speed X 1.0 Playback speed factor
--azimuth DEG 135 Camera yaw relative to the hit direction
--elevation DEG -15 Camera elevation
--distance M 3.2 Camera distance to the pelvis
--no-markers off Hide the spheres and the arrow
--record FILE off Render to .mp4 or .gif instead of opening a window
--every N 1 When recording, keep every N-th frame
--width, --height 1280, 720 Recording size in pixels
--scene XML robot/scene_tennis.xml Another MuJoCo model with the same 29 joint names
--root DIR this dataset Dataset folder

With --scene pointing at a stock G1 model the motion still plays; the yellow marker needs the racket_sweetspot site and is skipped when the model has none.

Use the loader in your own code

import sys
sys.path.insert(0, "scripts")
from fivehundredshots import JOINT_NAMES, load_clip, load_metadata

shots = load_metadata()                      # list of dicts, one per shot
clip = load_clip(shots[0]["file"])           # named (T, ...) arrays
print(clip.joint_pos.shape, JOINT_NAMES[0])  # (176, 29) left_hip_pitch_joint

MuJoCo orders the joints differently from the dataset, so place them by name:

qadr = [model.joint(name).qposadr[0] for name in JOINT_NAMES]
data.qpos[:3], data.qpos[3:7], data.qpos[qadr] = clip.root_pos[t], clip.root_quat[t], clip.joint_pos[t]