Cleaning Task Training Data

Claru collects cleaning task data across real homes, offices, and commercial spaces using mobile manipulators and table-mounted arms equipped with 100 Hz force/torque sensors. Our standardized mess creation protocol ensures reproducible diversity across liquid stains, semi-viscous substances, dried residue, and scattered clutter. Each episode includes before/after overhead images for automated cleanliness scoring, force profiles per cleaning stroke, coverage maps, and human-rated cleanliness assessments. We support surface wiping, floor mopping, scrubbing, and clutter organization across 10+ surface materials. Datasets are delivered with full force sensor calibration, surface material labels, cleaning tool metadata, and stratified splits by environment type, surface material, and mess category.

Robotic cleaning encompasses contact-rich manipulation tasks — wiping counters, scrubbing surfaces, mopping floors, vacuuming, and organizing clutter — that require force control, coverage planning, and cleanliness assessment. Unlike autonomous vacuum robots that follow simple coverage algorithms, manipulation-based cleaning requires understanding what is dirty, choosing the right tool and technique, applying appropriate force without damaging surfaces, and verifying the result. A robot wiping a kitchen counter must distinguish between a water stain (light wipe), dried food (scrubbing with force), and a delicate surface (gentle pressure) — all from visual observation alone.

The market pull is enormous. The global professional cleaning services market exceeds $340B annually, with labor shortages driving automation interest. Companies like Aeolus Robotics, Avidbots, and Hello Robot are building cleaning-capable mobile manipulators, but the data bottleneck is severe. TidyBot (Wu et al., 2023) demonstrated personalized tidying from a handful of demonstrations per household, but surface cleaning — which involves contact, force modulation, and quality assessment — requires 10-100x more demonstrations than non-contact rearrangement tasks.

The data challenge is three-fold. First, cleaning is inherently subjective — different people have different standards for 'clean enough.' This requires collecting demonstrations from multiple operators and annotating cleanliness on a continuous scale rather than a binary pass/fail. Second, cleaning involves sustained contact with force modulation, making force/torque sensing essential alongside vision. Third, coverage planning (ensuring every part of a surface is cleaned) requires trajectory annotations that go beyond point-to-point movements to capture sweeping, circular, and back-and-forth cleaning patterns.

Unlike most manipulation tasks where success is binary (the object is in the right place or it is not), cleaning success exists on a continuum. A surface can be partially clean, mostly clean, or spotlessly clean — and the acceptable threshold varies by deployment context. A hospital operating room demands sterilization-grade coverage, while a household kitchen counter just needs visible debris removed. This continuous quality spectrum means cleaning datasets must include not just demonstrations but paired quality assessments at multiple thresholds, enabling policies that adapt cleaning effort to the deployment standard. No other manipulation domain has this quality-adaptive requirement.

Research on learned cleaning policies lags behind other manipulation tasks because cleaning requires sustained contact — most robot learning research focuses on discrete pick-and-place or short-horizon tasks. The most relevant work comes from three directions. TidyBot (Wu et al., 2023) uses LLM-based personalization to learn tidying preferences from few-shot demonstrations, achieving 85% success on object placement but not addressing surface cleaning. Mobile ALOHA (Fu et al., 2024) demonstrated bimanual wiping and mopping from 50 demonstrations per task, with the bimanual setup enabling one arm to hold a cleaning tool while the other stabilizes the surface or object being cleaned.

Force-controlled wiping has been studied more extensively in industrial contexts. Luo et al. (2023) trained a hybrid force-position controller from demonstrations that adapts wiping force to the surface compliance — pressing harder on rigid counters and softer on wood. Their approach requires 500+ demonstrations per surface type to learn the force-compliance mapping. For coverage planning, Gaussian process models trained on human wiping trajectories can predict effective coverage paths, but these require dense trajectory annotations with per-timestep force readings — a data collection burden that limits scalability.

The frontier is whole-task cleaning: a robot that enters a room, assesses what needs cleaning, selects tools, cleans surfaces, organizes clutter, and verifies the result. This requires integrating navigation, manipulation, force control, and quality assessment in a single system. No existing dataset covers this full pipeline. Current approaches decompose it into subtask policies (navigate, pick tool, wipe, assess), each trained on separate datasets, with an LLM-based task planner orchestrating the sequence.

An emerging research direction uses visual language models to assess cleanliness from images. Rather than relying solely on before/after difference maps, these systems use VLMs to answer questions like 'is this counter clean enough for food preparation?' This semantic assessment aligns better with real deployment needs than pixel-level metrics. However, training these assessment models requires large datasets of images labeled with cleanliness judgments at multiple semantic levels — another data requirement that compounds the already demanding modality stack for cleaning robot training.

Cleaning data collection stations combine a mobile manipulator or table-mounted arm with surface workspaces of varying materials (granite, wood, stainless steel, glass, laminate), a set of cleaning tools (sponges, microfiber cloths, mop heads, spray bottles), and standardized mess creation protocols. Messes are created using reproducible methods: food coloring drops for liquid stains, ketchup smears for semi-viscous substances, dried coffee rings for aged stains, and scattered objects for clutter. This ensures diversity while maintaining reproducibility across collection sessions.

The annotation pipeline for cleaning data captures before-and-after cleanliness scores (rated on a 1-5 scale by human evaluators from overhead images), surface coverage percentage (computed from registered before/after frames using color difference thresholding), force profiles throughout the cleaning motion, and cleaning technique classification (linear wiping, circular scrubbing, spot treatment, sweeping). Episodes where the operator damages the surface (scratching, excessive water) are annotated as negative examples with damage type labels.

Environment diversity is critical for deployment generalization. Claru collects cleaning data across home kitchens, offices, bathrooms, and commercial spaces. Each environment introduces different surfaces, mess types, and spatial constraints. We use a before/after photography protocol where overhead images are captured pre- and post-cleaning under consistent lighting, enabling automated cleanliness assessment that supplements human ratings. A minimum of 5 environment types and 10 surface materials per collection campaign ensures policy robustness.

Operator fatigue management is especially important for cleaning data because the physical effort of sustained wiping and scrubbing degrades demonstration quality faster than other manipulation tasks. Sessions are limited to 25 minutes for scrubbing tasks (versus 45 minutes for standard teleoperation) with mandatory 15-minute breaks. Force profile analysis provides an objective fatigue indicator: when the operator's average wiping force drops below 70% of their session-start baseline, the session is terminated and the remaining demonstrations are flagged for quality review. This ensures that late-session demonstrations do not train the policy to under-clean.