Mouse Reaching Imitation Eval

---
name: mouse-reaching-imitation-eval
description: Evaluates an imitation learning agent's ability to reproduce mouse forelimb reaching kinematics and muscle activation patterns using a musculoskeletal physics model. It probes how well learned policies can match biological motion and electrophysiological signals under varying physics-aware constraints. Use when the user wants to benchmark on Mouse forelimb reaching mocap dataset, or asks about evaluating this task. Reports track replay error.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2511.21848
  bibtex_key: leonardis2025massively
  confidence: high
---

# mouse-reaching-imitation-eval

> Massively Parallel Imitation Learning of Mouse Forelimb Musculoskeletal Reaching Dynamics — Leonardis et al. (2025) (arXiv:2511.21848, 2025)

## What this evaluates

Evaluates an imitation learning agent's ability to reproduce mouse forelimb reaching kinematics and muscle activation patterns using a musculoskeletal physics model. It probes how well learned policies can match biological motion and electrophysiological signals under varying physics-aware constraints.

## Datasets

- **Mouse forelimb reaching mocap dataset** — total 46; splits: test (46)

## Metrics

- `track replay error` **(primary)** — range: mm
  - Mean absolute error (in mm) between the reference pose/joint angles and the imitated pose generated by the agent.
- `EMG fit MAE` — range: other
  - Mean absolute error between recorded in vivo EMG signals and simulated muscle activations.
- `Simplex $
ho$` — range: [-1, 1]
  - Pearson correlation coefficient from simplex projection used to decode simulated actions or predict EMG from joint angles.

## Input / output format

**Input**: 3D mocap keypoints and joint angles from recorded mouse reaching behavior, along with reference EMG signals.

**Output**: Simulated joint angles, muscle activations, and control actions from the imitation learning agent.

## Scoring recipe

```python
def compute_metrics(predictions, gold):
    kinematic_mae = np.mean(np.abs(predictions['joint_angles'] - gold['joint_angles']))
    emg_mae = np.mean(np.abs(predictions['muscle_activations'] - gold['emg']))
    simplex_rho = pearsonr(simplex_project(predictions['joint_angles']), gold['emg'])[0]
    return {'track_replay_error': kinematic_mae, 'emg_fit_mae': emg_mae, 'simplex_rho': simplex_rho}
```

## Common pitfalls

- High joint reward does not necessarily correlate with accurate EMG fitting; adding control or energy costs improves EMG fit but may decrease joint reward.
- Track replay error alone (<1mm) does not guarantee physiologically plausible muscle activations without physics-aware constraints.

## Evidence (verbatim from paper)

> The imitation learning performance (track replay) shows that in absence of physics aware constraints it can recapitulate the reference kinematics with high precision of less than 1mm on average... Adding control cost leads to a better fit between simulated and reference EMG for the biceps... The results show that Simplex projections of the joint angles over time can be used to decode the simulated actions, with simplex $
ho$ at .802 for Biceps.

## Citation

```bibtex
@misc{leonardis2025massively,
  title={Massively Parallel Imitation Learning of Mouse Forelimb Musculoskeletal Reaching Dynamics},
  author={Leonardis et al. (2025)},
  year={2025},
  note={arXiv:2511.21848}
}
```

- arXiv: 2511.21848