MimicDodo

Figure 1: Dodo Daimao Backflip Reference Trajectory

Figure 2: Dodo robot performing a backflip trained via RSL-RL

A unified workspace for Dodo robot motion generation and training, combining trajectory optimization with reinforcement learning-based whole-body tracking.

🧩 Prerequisites

• Ubuntu 22.04 (Linux x64)
• RAM: 32GB
• GPU VRAM: 16GB (recommended for training)
• NVIDIA GPU with compatible driver
• Conda (recommended)
• Isaac Sim 4.5.0 or later (for whole_body_tracking)

---

📁 Project Structure

MimicDodo/
├── assets/
│   ├── dodo/                          # Original Dodo robot
│   │   ├── urdf/dodo.urdf
│   │   ├── meshes/
│   │   └── config/
│   └── dodo_daimao/                   # DodoDaimao robot
│       ├── urdf/dodo_daimao.urdf      # Fixed: hip_left mass, primitive collisions
│       └── meshes/
│
├── trajectory_creation/
│   ├── pinocchio/                     # Simple FK/IK recording scripts
│   └── se3_trajopt/src/
│       ├── examples/agile_exps/
│       │   ├── dodo_backflip.py
│       │   ├── dodo_daimao_backflip.py
│       │   └── dodo_daimao_walk.py
│       └── robots/
│           ├── dodobot_v3/DodoWrapper.py
│           └── dodo_daimao/DodoDaimaoWrapper.py
│
└── whole_body_tracking/
    ├── dodo_daimao_backflip.npz       # Generated — do not commit
    ├── scripts/
    │   ├── csv_to_npz_dodo.py         # Dodo converter (requires Isaac Sim)
    │   ├── csv_to_npz_dodo_daimao.py  # DodoDaimao converter (Pinocchio, fast)
    │   ├── replay_npz_dodo.py
    │   └── rsl_rl/
    │       ├── train.py
    │       └── play.py
    └── source/whole_body_tracking/whole_body_tracking/
        ├── robots/
        │   ├── dodo.py
        │   └── dodo_daimao.py
        └── tasks/tracking/
            ├── dodo/
            │   ├── __init__.py
            │   ├── flat_env_cfg.py
            │   └── agents/rsl_rl_ppo_cfg.py
            └── dodo_daimao/
                ├── __init__.py
                ├── flat_env_cfg.py
                └── agents/
                    ├── __init__.py
                    └── rsl_rl_ppo_cfg.py

---

🚀 Installation

Trajectory Creation (Pinocchio / SE3 TrajOpt)

conda create -n trajopt python=3.13
conda activate trajopt
conda install -c conda-forge pinocchio meshcat-python cyipopt matplotlib numpy

Whole Body Tracking (Isaac Lab)

cd whole_body_tracking
conda create -n beyondmimic python=3.10
conda activate beyondmimic
pip install --upgrade pip
pip install "isaacsim[all,extscache]==4.5.0.0" --extra-index-url https://pypi.nvidia.com
pip install -U torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu124
./isaaclab.sh --install rsl_rl
pip install wandb
python -m pip install -e source/whole_body_tracking

---

🎬 DodoDaimao Backflip — Full Workflow

All commands run from the MimicDodo project root unless noted.

Step 1 — Generate backflip trajectory

conda activate trajopt
export PYTHONPATH="$(pwd)/trajectory_creation/se3_trajopt/src"

# With visualizer (meshcat — open http://127.0.0.1:7000/static/ in browser)
python trajectory_creation/se3_trajopt/src/examples/agile_exps/dodo_daimao_backflip.py --vis

# Headless export only
python trajectory_creation/se3_trajopt/src/examples/agile_exps/dodo_daimao_backflip.py

# Output: whole_body_tracking/scripts/dodo_daimao_backflip.csv

Step 2 — Convert CSV → NPZ (fast, no Isaac Sim needed)

# Still in trajopt env — finishes in <5 seconds
python whole_body_tracking/scripts/csv_to_npz_dodo_daimao.py \
  --input_file  whole_body_tracking/scripts/dodo_daimao_backflip.csv \
  --input_fps   100 \
  --output_name dodo_daimao_backflip \
  --output_fps  100

# Output: whole_body_tracking/dodo_daimao_backflip.npz

Step 3 — Upload NPZ to wandb

cd whole_body_tracking
conda activate beyondmimic

python -c "
import wandb
name = 'dodo_daimao_backflip'
run  = wandb.init(project='csv_to_npz', name=name)
art  = run.log_artifact('dodo_daimao_backflip.npz', name=name, type='motions')
run.link_artifact(art, target_path=f'wandb-registry-motions/{name}')
print('Uploaded: motions/' + name)
run.finish()
"

Step 4 — Replay to verify (recommended before training)

# From whole_body_tracking/
python scripts/replay_npz_dodo.py \
  --registry_name simon-kruelle-technical-university-of-munich/csv_to_npz/dodo_daimao_backflip

Check that the motion looks correct in Isaac Sim before spending GPU time training.

Step 5 — Train

# From whole_body_tracking/
WANDB_DIR=/tmp/wandb_cache python scripts/rsl_rl/train.py \
  --task=Tracking-Flat-DodoDaimao-v0 \
  --motion_file dodo_daimao_backflip.npz \
  --wandb_project whole_body_tracking \
  --headless

To resume from a checkpoint:

WANDB_DIR=/tmp/wandb_cache python scripts/rsl_rl/train.py \
  --task=Tracking-Flat-DodoDaimao-v0 \
  --motion_file dodo_daimao_backflip.npz \
  --wandb_project whole_body_tracking \
  --resume=True \
  --load_run=<your-run-id> \
  --checkpoint=model_XXXXX.pt \
  --headless

Step 6 — Evaluate trained policy

# From whole_body_tracking/
python scripts/rsl_rl/play.py \
  --task=Tracking-Flat-DodoDaimao-v0 \
  --num_envs=1 \
  --load_run "<your-run-id>" \
  --checkpoint "model_XXXXX.pt" \
  --motion_file dodo_daimao_backflip.npz \
  --video \
  --video_length 1000 \
  --disable_fabric

---

🎬 Original Dodo Backflip — Full Workflow

Step 1 — Generate trajectory

conda activate trajopt
export PYTHONPATH="$(pwd)/trajectory_creation/se3_trajopt/src"
python trajectory_creation/se3_trajopt/src/examples/agile_exps/dodo_backflip.py --vis

Step 2 — Convert CSV → NPZ (requires Isaac Sim)

cd whole_body_tracking
conda activate beyondmimic
python scripts/csv_to_npz_dodo.py \
  --input_file scripts/dodo_backflip.csv \
  --input_fps 100 \
  --output_name dodo_backflip \
  --output_fps 100

Step 3 — Train

WANDB_DIR=/tmp/wandb_cache python scripts/rsl_rl/train.py \
  --task=Tracking-Flat-Dodo-v0 \
  --motion_file dodo_backflip.npz \
  --wandb_project whole_body_tracking \
  --headless

Step 4 — Evaluate

python scripts/rsl_rl/play.py \
  --task=Tracking-Flat-Dodo-v0 \
  --num_envs=1 \
  --load_run "<your-run-id>" \
  --checkpoint "model_XXXXX.pt" \
  --motion_file dodo_backflip.npz \
  --video \
  --video_length 1000 \
  --disable_fabric

---

🎬 Pinocchio Recording Scripts (Simple Motions)

conda activate trajopt
export PYTHONPATH="$(pwd)/trajectory_creation/se3_trajopt/src"

python trajectory_creation/pinocchio/record_walk.py
python trajectory_creation/pinocchio/record_jump.py
python trajectory_creation/pinocchio/record_kangaroo.py
python trajectory_creation/pinocchio/record_stand.py

Output: CSV files saved to whole_body_tracking/scripts/

---

🤖 Robot Configurations

Dodo (original)

• URDF: assets/dodo/urdf/dodo.urdf
• Joints: left_joint_1–4, right_joint_1–4
• Isaac Lab task: Tracking-Flat-Dodo-v0
• CSV converter: scripts/csv_to_npz_dodo.py (requires Isaac Sim)

DodoDaimao

• URDF: assets/dodo_daimao/urdf/dodo_daimao.urdf
• Joints: hip_left/right, upper_leg_left/right, lower_leg_left/right, foot_left/right
• Isaac Lab task: Tracking-Flat-DodoDaimao-v0
• CSV converter: scripts/csv_to_npz_dodo_daimao.py (Pinocchio-based, fast)
• Key fixes applied:
  • hip_left mass corrected 0 → 0.4286 kg (was causing asymmetric dynamics)
  • All collision geometries replaced with primitives (box for body and feet, none for legs/hips) — matches dodo.urdf pattern and is required for Isaac Sim performance
  • Two contact points per foot (toe + heel) added in wrapper — required for yaw stability
  • Standing height corrected to q[2] = 0.435902 (feet exactly at z=0)

---

📋 Quick Reference Cheat Sheet

DodoDaimao

# From MimicDodo root
conda activate trajopt
export PYTHONPATH="$(pwd)/trajectory_creation/se3_trajopt/src"
python trajectory_creation/se3_trajopt/src/examples/agile_exps/dodo_daimao_backflip.py
python whole_body_tracking/scripts/csv_to_npz_dodo_daimao.py --input_file whole_body_tracking/scripts/dodo_daimao_backflip.csv --input_fps 100 --output_name dodo_daimao_backflip --output_fps 100

cd whole_body_tracking && conda activate beyondmimic
python -c "import wandb; run=wandb.init(project='csv_to_npz',name='dodo_daimao_backflip'); art=run.log_artifact('dodo_daimao_backflip.npz',name='dodo_daimao_backflip',type='motions'); run.link_artifact(art,target_path='wandb-registry-motions/dodo_daimao_backflip'); run.finish()"
python scripts/replay_npz_dodo.py --registry_name simon-kruelle-technical-university-of-munich/csv_to_npz/dodo_daimao_backflip
WANDB_DIR=/tmp/wandb_cache python scripts/rsl_rl/train.py --task=Tracking-Flat-DodoDaimao-v0 --motion_file dodo_daimao_backflip.npz --wandb_project whole_body_tracking --headless

Dodo

conda activate trajopt && export PYTHONPATH="$(pwd)/trajectory_creation/se3_trajopt/src"
python trajectory_creation/se3_trajopt/src/examples/agile_exps/dodo_backflip.py

cd whole_body_tracking && conda activate beyondmimic
python scripts/csv_to_npz_dodo.py --input_file scripts/dodo_backflip.csv --input_fps 100 --output_name dodo_backflip --output_fps 100
WANDB_DIR=/tmp/wandb_cache python scripts/rsl_rl/train.py --task=Tracking-Flat-Dodo-v0 --motion_file dodo_backflip.npz --wandb_project whole_body_tracking --headless

---

🛠️ Troubleshooting

Isaac Sim loads very slowly / freezes on URDF import

All <collision> bodies in the URDF must use primitive shapes (<box>, <cylinder>, <sphere>), not <mesh>. Full mesh collision causes Isaac Sim's convex decomposition to run on every triangle. After editing the URDF, clear the USD cache so Isaac Sim regenerates it:

find ~/.local/share/ov -name "*dodo_daimao*" -delete 2>/dev/null
find /tmp -name "*dodo_daimao*" -delete 2>/dev/null

Also ensure force_usd_conversion=True in dodo_daimao.py.

csv_to_npz is slow (2–3 min)

Use csv_to_npz_dodo_daimao.py instead of csv_to_npz_dodo.py. The dodo_daimao version uses Pinocchio FK and runs in the trajopt conda env in under 5 seconds — no Isaac Sim boot required.

Yaw rotation / spinning during walking or backflip

The DodoDaimaoWrapper must register two contact frames per foot (toe + heel). A single point contact has no yaw resistance and the solver freely spins the body. Check DodoDaimaoWrapper.__init__ for the _add_foot_frame calls.

Wrong foot height / feet below ground at start

go_neutral() uses hardcoded q[2] = 0.435902. If you change the neutral joint angles, recompute with:

robot = DodoDaimao()
q = robot.go_neutral()
robot.fk_all(q)
fid = robot.model.getFrameId("left_toe")
foot_z = robot.data.oMf[fid].translation[2]
print(foot_z)  # should be ~0.0
# If not: set q[2] = 0.435902 - foot_z and hardcode the result

k1 == k2 assertion error in backflip script

The flight phase must produce at least 1 timestep: flight_duration / DT >= 1. Current setting: 0.45s / 0.05s = 9 steps. If you change DT, scale the flight duration accordingly.

hip_left zero mass (asymmetric dynamics)

Already fixed in the current dodo_daimao.urdf. If you regenerate the URDF from SolidWorks, check that hip_left has <mass value="0.428645410212592"/> — the exporter originally set it to 0.

---

📄 License

See individual project directories for license information.