Deepmind Control Suite Eval

---
name: deepmind-control-suite-eval
description: Evaluates continuous control reinforcement learning agents on a standardized suite of physics-based simulation tasks. It probes sample efficiency, stability, and performance over long training horizons using uniform action, observation, and reward structures. Use when the user wants to benchmark on DeepMind Control Suite, or asks about evaluating this task. Reports return.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1801.00690
  bibtex_key: tassa2018deepmind
  confidence: high
---

# deepmind-control-suite-eval

> DeepMind Control Suite — Tassa et al. (2018) (arXiv:1801.00690, 2018)

## What this evaluates

Evaluates continuous control reinforcement learning agents on a standardized suite of physics-based simulation tasks. It probes sample efficiency, stability, and performance over long training horizons using uniform action, observation, and reward structures.

## Datasets

- **DeepMind Control Suite** — total ?; splits: tasks (-1)

## Metrics

- `return` **(primary)** — range: varies by task
  - Sum of rewards collected over a single episode. Performance is aggregated as the median, 5th percentile, and 95th percentile across multiple runs.

## Input / output format

**Input**: Low-dimensional state feature vectors from the MuJoCo physics simulation environment.

**Output**: Continuous action vectors sampled from a standardized action space.

## Scoring recipe

```python
all_returns = []
for seed in range(15):
    for step in range(0, 1e8, 100000):
        for _ in range(10):
            ret = run_episode(exploration_noise=False)
            all_returns.append(ret)
median_ret = median(all_returns)
p5_ret = percentile(all_returns, 5)
p95_ret = percentile(all_returns, 95)
```

## Common pitfalls

- Using exploration noise during evaluation (protocol explicitly specifies no exploration noise)
- Not aggregating results over the full 1e8 step horizon
- Failing to report median and 5th/95th percentiles across the required 15 seeds per task

## Evidence (verbatim from paper)

> The plots in Figure 4 and Figure 5 show the median and the 5th and 95th percentile of the returns for the first 1e8 steps. Each agent was run 15 times per task using different seeds (except for D4PG which was run 5 times), using only low-dimensional state feature information.

## Citation

```bibtex
@misc{tassa2018deepmind,
  title={DeepMind Control Suite},
  author={Tassa et al. (2018)},
  year={2018},
  note={arXiv:1801.00690}
}
```

- arXiv: 1801.00690