robo-augment

Online image augmentation toolkit for robot learning. 44 methods in 9 functional groups, designed for visuomotor policy training.

Why

Every major robot learning framework uses at most 3 augmentations. RAD (NeurIPS 2020) showed random crop alone gives 5.75x improvement, yet the field stopped there. No pip-installable, online, robotics-aware augmentation library exists.

robo-augment fills this gap with 44 transforms organized into 9 functional groups, per-camera intensity scaling, augmentation symmetry verification, and speed-optimized pure-PyTorch implementation.

See docs/motivation.md for detailed research motivation and per-method evidence.

Install

pip install robo-augment

Quick Start

from roboaug import create_default_pipeline

aug = create_default_pipeline(camera_role="top", scale=0.6)
augmented = aug(image_tensor)  # (C, H, W) float32 [0, 1]

Presets

from roboaug.presets import manipulation_preset, minimal_preset, sim2real_preset, navigation_preset

# Full 44-method pipeline for tabletop manipulation
aug = manipulation_preset()

# Multi-camera with per-camera scaling
augs = manipulation_preset(cameras={"top": 0.6, "wrist": 0.85, "side": 1.0})

# DrQ-style minimal (color jitter + crop + noise)
aug = minimal_preset()

# Heavy photometric for sim-to-real transfer
aug = sim2real_preset()

# Navigation-safe (no spatial transforms that break heading)
aug = navigation_preset()

Custom Pipeline

from roboaug import GroupedAugment
from roboaug.transforms import GaussianNoise, MotionBlur, RandomShadow

pipeline = GroupedAugment(groups=[
    {"name": "noise", "mode": "exclusive", "p": 0.3, "transforms": [
        (GaussianNoise(std=(5, 25)), 3.0),
        (MotionBlur(kernel_size=(3, 11)), 2.0),
    ]},
    {"name": "lighting", "mode": "exclusive", "p": 0.3, "transforms": [
        (RandomShadow(opacity=(0.2, 0.5)), 1.0),
    ]},
], camera_scale=0.6)

augmented = pipeline(image_tensor)

9 Functional Groups

Independent mode: each transform fires with its own probability, multiple can activate per sample.

Exclusive mode: at most one transform fires per sample from the group, preventing unrealistic compounding.

Framework Integration

LeRobot

from roboaug.presets import manipulation_preset
from lerobot.datasets.lerobot_dataset import LeRobotDataset

augs = manipulation_preset(cameras={"top": 0.6, "wrist": 0.85, "side": 1.0})
dataset = LeRobotDataset(..., image_transforms=augs)

Diffusion Policy

from roboaug import create_default_pipeline

aug = create_default_pipeline(camera_role="side")
# Apply in your dataset __getitem__:
obs["image"] = aug(obs["image"])

Design Principles

1. Action-aware: spatial transforms bounded to preserve action semantics (max 10° rotation, 10% translation). No horizontal flip (breaks left-right actions).
2. Per-camera scaling: different cameras tolerate different augmentation intensity. Top overview cameras get lighter aug than side context cameras.
3. Symmetric augmentation: all transforms verified to not shift BatchNorm statistics (< 0.005 mean bias). Shadow, gradient, gamma, vignette all have symmetric brighten/darken modes.
4. Pure PyTorch: zero cv2 dependency in the forward path. 44 methods at ~14ms per 480x640 image.
5. Grouped sampling: functional groups prevent unrealistic effect compounding while maximizing diversity.

Speed

Benchmarked on Intel Xeon (single core, CPU):

Citation

@software{roboaugment2026,
    title={robo-augment: Online Image Augmentation for Robot Learning},
    author={Li, Yuxian},
    year={2026},
    url={https://github.com/Liyux3/robo-augment}
}

License

Apache 2.0