Dataset - BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion

Dataset

Content

dataset_root/
├─ rosbag_agile_motion/
│  ├─ C1970_tkd_skill_clip1.mcap
│  ├─ C1975_side_flip.mcap
│  ├─ C1980_double_high_kick.mcap
│  └─ C1985_merge2.mcap   # side_flip + tkd_skill
│
├─ rosbag_walk_and_run/
│  ├─ walk_rosbag
│  └─ run_rosbag
│
└─ rosbag_ablation/
   ├─ 2ms_rosbag2_2025_10_22-03_50_34
   ├─ 5ms_rosbag2_2025_10_22-03_41_01
   ├─ 10ms_rosbag2_2025_10_22-03_36_03
   ├─ arma0_rosbag2_2025_10_22-00_29_58
   ├─ arma0.1_rosbag2_2025_10_22-00_44_01
   ├─ arma10_rosbag2_2025_10_22-00_48_58
   ├─ asapgain_rosbag2_2025_10_22-01_03_25
   ├─ axisangle_rosbag2_2025_10_22-01_24_25
   ├─ hist4_rosbag2_2025_10_22-07_00_11
   ├─ hist8_rosbag2_2025_10_22-07_03_51
   ├─ hist25_rosbag2_2025_10_22-07_06_58
   ├─ origin_rosbag2_2025_10_22-01_35_46
   ├─ quat_rosbag2_2025_10_22-01_17_38
   ├─ wn5_rosbag2_2025_10_22-00_54_05
   └─ wn25_rosbag2_2025_10_22-00_57_52

1) rosbag_agile_motion/

Agile acrobatic clips recorded on the humanoid:

• C1970_tkd_skill_clip1.mcap — Taekwondo skill clip.
• C1975_side_flip.mcap — Side-flip sequence.
• C1980_double_high_kick.mcap — Double high-kick sequence.
• C1985_merge2.mcap — Composite sequence (side_flip + tkd_skill).

2) rosbag_walk_and_run/

• walk_rosbag — Walking trials.
• run_rosbag — Running/sprinting trials.

3) rosbag_ablation/

Ablation study recordings:

Code for Plot

1. IMU data plot

Directory: ./base_imu/

We first use plotjuggler to export the .csv file from .mcap bag 

python export_imu_csv.py rosbag_data_raw.csv
python plot_imu_csv.py

2. Plot local motion tracking err

Directory: ./ablation/

2.1. Plot raw joint position vs ref motion joint position data

python plot_joint_raw.py

Please change the `BAG` path in the code.

2.2. Batch run for local err
Plot local motion tracking err for all the bags under `BASE_DIR`.

Please change the `BASE_DIR` in the code.

python batch_local.py

The results will be stored at the path of the dataset. 

3. Plot global motion tracking err

3.1. Batch run for global err
Plot global motion tracking err for all the bags under `BASE_DIR`.

Please change the `BASE_DIR` in the code.

python batch_global.py

Visualize global err of all the experiments: Use Chrome open: `mocap_error_gallery_local_folder_viewer.html`

4. Plot overall err

Compare local and global err among selected experiments.

Example:
python compare_err.py --base "/media/nv/TSD302/rosbag_ablation" --exp origin hist4 hist8 hist25

Batch run for overall err
Compare the global and local motion tracking err among experiments.

Please change the `BASE_DIR` in the code.

python batch_all_err.py

5. Adaptive Sampling Plot

5.1. Plot Adaptive Sampling Failure Hatmap

python plot_failure_heatmap.py   --dir ./w \                                    
  --out_dir ./w_out \
  --bins 200 \
  --thresh 0.05 \
  --dpi 180 \
  --pattern "*_ckpt_failure_*.npy" --save_csv

5.2. Plot Adaptive Sampling Prob over Time

python plot_sampling_prob.py sampling_prob_over_time.pkl

6. GRF Plot

6.1. Walking

Human ref:
python grf_walk_ref.py

Robot real:
python run_bag.py --csv walk_bag2.csv --out walk_bag2.png

6.2. Running

Human ref:
python grf_run_ref.py RBDS028runT25forces.txt --save ref_run_grf.png

Robot real:
python run_bag.py --csv run_bag.csv --out run_bag.png