EpicPinkPenguin/visual_distracting_control_suite

Visual Distracting Control Suite Benchmark

This dataset contains expert trajectories generated by a Proximal Policy Optimization (PPO) reinforcement learning agent trained on 4 environments of the Distracting Control Suite. For each environment we collect data with different levels of distraction, which we define below, and masks for the agent.

Levels of distraction:

• None: Vanilla DeepMind Control Suite without visual distractions. The environment uses the default static background and maintains a fixed camera position.
• Low: Dynamic background distractors only.
  • Background: Video frames from the DAVIS dataset are played sequentially. At each step, the background advances to the next frame, cycling through the video (direction reverses at endpoints).
  • Camera: Fixed position, no scale or rotation changes (scale=0.0).
• Hard: Dynamic background and camera distractors.
  • Background: Same dynamic video background as "Low" level.
  • Camera: Dynamic camera during an episode, changes position, rotation, and zoom (scale=0.1).
    • Horizontal/vertical rotation delta: ±π/20 radians (±9°)
    • Roll rotation delta: ±π/20 radians (±9°)
    • Zoom range: 5% zoom in to 15% zoom out relative to default distance
    • Dynamic velocity: position velocity std=0.01, max velocity=0.04
    • Dynamic roll: roll std=π/3000, max roll velocity=π/500

Dataset Usage

Regular usage (for the domain acrobot with task swingup):

from datasets import load_dataset
train_dataset = load_dataset("EpicPinkPenguin/visual_distracting_control_suite", name="cheetah_run_distractor_hard", split="train")
test_dataset = load_dataset("EpicPinkPenguin/visual_distracting_control_suite", name="cheetah_run_distractor_hard", split="test")

Agent Performance

The PPO agent was trained for 2M steps on each environment and obtained the following final performance metrics on the evaluation environment.

Environment	Steps (Train)	Steps (Test)	Return	Observation
cheetah_run	9,000,000	1,000,000	837.67
cheetah_run_distractor_low	9,000,000	1,000,000	837.67
cheetah_run_distractor_hard	9,000,000	1,000,000	837.67
hopper_hop	9,000,000	1,000,000	307.33
hopper_hop_distractor_low	9,000,000	1,000,000	307.33
hopper_hop_distractor_hard	9,000,000	1,000,000	307.33
humanoid_walk	9,000,000	1,000,000	616.52
humanoid_walk_distractor_low	9,000,000	1,000,000	616.52
humanoid_walk_distractor_hard	9,000,000	1,000,000	616.52
walker_run	9,000,000	1,000,000	738.37
walker_run_distractor_low	9,000,000	1,000,000	738.37
walker_run_distractor_hard	9,000,000	1,000,000	738.37

Dataset Structure

Data Instances

Each data instance represents a single step consisting of tuples of the form (observation, state, mask, action, reward, done, truncated) = (o_t, s_t, m_t, a_t, r_t, terminated_t, truncated_t).

{'action': [1],
 'observation': [[[0, 166, 253],
                  [0, 174, 255],
                  [0, 170, 251],
                  [0, 191, 255],
                  [0, 191, 255],
                  [0, 221, 255],
                  [0, 243, 255],
                  [0, 248, 255],
                  [0, 243, 255],
                  [10, 239, 255],
                  [25, 255, 255],
                  [0, 241, 255],
                  [0, 235, 255],
                  [17, 240, 255],
                  [10, 243, 255],
                  [27, 253, 255],
                  [39, 255, 255],
                  [58, 255, 255],
                  [85, 255, 255],
                  [111, 255, 255],
                  [135, 255, 255],
                  [151, 255, 255],
                  [173, 255, 255],
...
                  [0, 0, 37],
                  [0, 0, 39]]],
 'state': [-0.09255199134349823, 0.028468089178204536, -0.05743644759058952, ..., -0.013366516679525375, -0.08739502727985382, 0.007727491203695536]
 'mask' = [
    [0,   0,   0,   0, ...,   0,   0,   0,   0],
    [0,   0,   0,   0, ...,   0,   0,   0,   0],
    [0,   0, 255, 255, ..., 255, 255,   0,   0],
    [0,   0, 255, 255, ..., 255, 255,   0,   0],
...
    [0,   0, 255, 255, ..., 255, 255,   0,   0],
    [0,   0, 255, 255, ..., 255, 255,   0,   0],
    [0,   0,   0,   0, ...,   0,   0,   0,   0],
    [0,   0,   0,   0, ...,   0,   0,   0,   0],
]
 'reward': 0.0,
 'terminated': False
 'truncated': False}

Data Fields

• observation: The current RGB observation from the environment.
• state: The current state of the environment.
• mask: A segmentation mask of the agent, with everything zero, except the agent, which is 255.
• action: The action predicted by the agent for the current observation.
• reward: The received reward for the current observation.
• terminated: If the episode has terminated with the current observation.
• truncated: If the episode is truncated with the current observation.

Data Splits

The dataset is divided into a train (90%) and test (10%) split. Each environment-dataset has in sum 10M steps (data points).

Dataset Creation

The dataset was created by training a PPO RL agent 2M steps in each environment. The trajectories where generated by taking a greedy action (mean) from the predicted action distribution at each step. The agent was trained on the state. Each environment was created with the same random seed, making the trajectories identical between the different distraction levels. This means concretely, that episode 0 of cheetah_run is identical with episode 0 of cheetah_run_distractor_low and cheetah_run_distractor_hard in everything, except the observation due to the visual distractors. This continues for the remaining episodes.

Distracting Control Suite

The Distracting Control Suite is an extension of the DeepMind Control Suite that augments standard continuous control tasks with visual distractions to evaluate the robustness of reinforcement learning (RL) algorithms. While preserving the underlying MuJoCo-based physics and task dynamics, it introduces changes in the visual observations—such as background videos, colors, textures, and camera variations—that are unrelated to the control objective. These distractions are designed to challenge agents’ ability to learn representations that generalize beyond spurious visual correlations. By decoupling task-relevant dynamics from high-dimensional, non-stationary visual noise, the Distracting Control Suite provides a controlled benchmark for studying generalization, representation learning, and robustness in vision-based RL. It is commonly used to assess how well algorithms trained in one visual setting transfer to others, and to compare methods that aim to improve invariance, stability, and sample efficiency under perceptual perturbations.