ampmove

Bipedal walking for iRonCub-Mk1_1 via two parallel tracks:

• Track 1 — PPO baseline: train from scratch with SB3 in MuJoCo
• Track 2+3 — AMP: retarget CMU mocap → motion prior → AMP training in Isaac Lab with SKRL

Based on the ADHERENT pipeline (joint-space retargeting, no IK).

---

Status

Step	Status
CMU BVH → iRonCub joint angles (retargeting)	done
MuJoCo visualization of retargeted motion	done
Convert .npy → Isaac Lab .npz (FK + 6D rot + contact)	done
PPO gymnasium env (MuJoCo)	done
PPO training + inference scripts	done
Isaac Lab simulation (motion replay)	done
Isaac Lab AMP task config	pending
SKRL AMP training	pending

---

Repo structure

ampmove/
├── assets/iRonCub/
│   ├── meshes/stl/                        # STL meshes (tracked in repo)
│   ├── meshes/obj/                        # OBJ meshes (tracked in repo)
│   └── robots/iRonCub-Mk1_1/
│       ├── model.urdf                     # iDynTree FK (retargeting)
│       ├── model_stl.urdf                 # MuJoCo + Isaac Lab USD conversion
│       └── ironcub.usd                   # generated per-machine (see Isaac Lab setup)
│
├── configs/
│   └── ironcub_bvh_mapping.yaml
│
├── motion_priors/walking/
│   ├── 07_12.bvh                          # source CMU mocap (walking, subject 07)
│   ├── 07_12_retargeted_adherent.npy      # retargeted motion (root pose + joint angles)
│   └── 07_12_isaaclab.npz                # Isaac Lab AMP input (FK link states + contacts)
│
├── retargeting/
│   ├── retarget_bvh_npy.py               # BVH → .npy
│   └── npy_to_npz.py                     # .npy → .npz (FK states + contacts for AMP)
│
├── mujoco_track/
│   ├── env.py                             # Gymnasium env (26-DOF, position control)
│   ├── visualize_retargeted.py            # motion replay in MuJoCo viewer
│   ├── train_ppo.py                       # SB3 PPO training
│   └── infer_ppo.py                       # load checkpoint + visualize
│
└── isaac_lab_track/
    ├── ironcub_cfg.py                     # ArticulationCfg
    ├── visualize_retargeted.py            # motion replay in Isaac Lab viewer
    └── (task.py, train_amp.py — pending)

---

Setup

1. Clone

git clone <repo-url>
cd ampmove

Meshes are tracked — no separate download needed.

2. Conda environment

conda create -n retarget python=3.10
conda activate retarget

3. Install idyntree with IPOPT

conda install -c robotology idyntree

Must be the conda version — pip idyntree does not include IPOPT.

4. Remaining dependencies

pip install mujoco==3.3.0 bvh scipy numpy stable-baselines3 gymnasium

---

Track 1 — PPO (MuJoCo)

# Train
python mujoco_track/train_ppo.py
python mujoco_track/train_ppo.py --timesteps 3000000 --n-envs 4

# Resume
python mujoco_track/train_ppo.py --resume logs/ppo_ironcub/best_model.zip

# Visualize policy
python mujoco_track/infer_ppo.py logs/ppo_ironcub/best_model.zip

# Visualize retargeted motion
python mujoco_track/visualize_retargeted.py

Checkpoints → checkpoints/ppo_ironcub/, best model → logs/ppo_ironcub/best_model.zip.

---

Track 2+3 — AMP (Isaac Lab)

Step 1 — Convert URDF to USD (once per machine)

Requires Isaac Lab installed. Run from repo root:

python {ISAACLAB}/scripts/tools/convert_urdf.py \
    assets/iRonCub/robots/iRonCub-Mk1_1/model_stl.urdf \
    assets/iRonCub/robots/iRonCub-Mk1_1/ironcub.usd \
    --merge-fixed-joints

ironcub.usd is generated locally and not tracked in git.

Step 2 — Simulate (motion replay, no RL)

python isaac_lab_track/visualize_retargeted.py
python isaac_lab_track/visualize_retargeted.py --headless

Replays 07_12_retargeted_adherent.npy at 120fps — good for verifying the asset and motion look correct before training.

Step 3 — AMP training

Pending.

---

Regenerating motion priors (already done, for reference)

# BVH → retargeted .npy
python retargeting/retarget_bvh_npy.py \
    motion_priors/walking/07_12.bvh \
    motion_priors/walking/07_12_retargeted_adherent.npy \
    configs/ironcub_bvh_mapping.yaml

# .npy → Isaac Lab .npz
python retargeting/npy_to_npz.py \
    motion_priors/walking/07_12_retargeted_adherent.npy \
    motion_priors/walking/07_12_isaaclab.npz

---

Retargeting — design notes

No IK. iRonCub has 26 DOFs. With only 2 foot position targets, IPOPT finds degenerate solutions. BVH Euler angles are mapped directly to robot joints using URDF axis analysis.

Coordinate transform. CMU BVH is Y-up, walk direction +Z. iRonCub world is Z-up, forward +X.

BVH2W = [[0,0,1], [1,0,0], [0,1,0]]

Sign conventions:

BVH channel	Robot joint	Sign	Reason
UpLeg Xrotation	hip_pitch	−1	URDF axis [0,−1,0]
Leg Xrotation	knee	−1	URDF axis [0,−1,0]
Foot Xrotation	ankle_pitch	+1	URDF axis [0,+1,0]

MuJoCo loading. Always use model_stl.urdf via MjSpec.from_file() with meshdir set. model.xml has an Rx(π) bug on leg bodies and is deleted.