Hri30 Action Surface Eval

---
name: hri30-action-surface-eval
description: Evaluates a vision-based framework for classifying object surfaces and recognizing human actions, and tests its integration into a human-robot collaboration controller for ergonomic task execution and subjective workload assessment. Use when the user wants to benchmark on HRI30, Custom Surface Dataset, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2207.03435
  bibtex_key: tassi2022sociable
  confidence: high
---

# hri30-action-surface-eval

> Sociable and Ergonomic Human-Robot Collaboration through Action Recognition and Augmented Hierarchical Quadratic Programming — Tassi et al. (2022) (arXiv:2207.03435, 2022)

## What this evaluates

Evaluates a vision-based framework for classifying object surfaces and recognizing human actions, and tests its integration into a human-robot collaboration controller for ergonomic task execution and subjective workload assessment.

## Datasets

- **HRI30** — total 2940; splits: train (-1), test (-1)
- **Custom Surface Dataset** — total ?; splits: train (-1), val (-1), test (-1)

## Metrics

- `accuracy` **(primary)** — range: [0, 1]
  - Number of correctly classified instances divided by the total number of instances in the evaluation set.
- `ergonomics_score ($e_s$)` — range: other
  - Proprietary ergonomic risk metric computed by the AHQP controller based on human pose and robot end-effector constraints; lower values indicate better posture and lower ergonomic risk.
- `NASA-TLX` — range: other
  - Subjective workload questionnaire measuring six subscales (temporal demand, mental demand, effort, physical demand, performance, frustration) on a 0-20 scale.

## Input / output format

**Input**: 224×224 pixel images for surface classification; 256×256 pixel video frames sampled at 1/4 rate for action recognition; real-time RGB-D camera feeds and human kinematic cues for HRC task execution.

**Output**: Binary surface classification label; action class label; robot end-effector trajectory and ergonomics score; NASA-TLX questionnaire responses.

## Scoring recipe

```python
def compute_accuracy(predictions, gold):
    correct = sum(1 for p, g in zip(predictions, gold) if p == g)
    return correct / len(gold)

def compute_ergonomics_score(human_pose, robot_ee_pose):
    # Proprietary AHQP internal metric
    return ahqp_controller.calculate_risk(human_pose, robot_ee_pose)

def compute_nasa_tlx(responses):
    # responses: dict of 6 subscales, each 0-20
    return {k: v for k, v in responses.items()}
```

## Common pitfalls

- Datasets are custom/internal and not publicly released, preventing direct replication or external benchmarking.
- Ergonomics score $e_s$ is a proprietary metric specific to the AHQP framework, not a standard biomechanical measure like REBA or RULA.
- NASA-TLX performance (PE) scale is inverted (higher score = better performance), which can cause misinterpretation if not explicitly noted.

## Evidence (verbatim from paper)

> The comparison shows that the baseline Resnet-152 outperforms the ResNet-50, given the number of layers that allow for more detailed feature extraction. Instead, the models based on our approach increased in accuracy by +6% for the ResNet-152 and by +10% for the Resnet-50 network, reaching 100% accuracy.

## Citation

```bibtex
@misc{tassi2022sociable,
  title={Sociable and Ergonomic Human-Robot Collaboration through Action Recognition and Augmented Hierarchical Quadratic Programming},
  author={Tassi et al. (2022)},
  year={2022},
  note={arXiv:2207.03435}
}
```

- arXiv: 2207.03435