History 1 Touch, then wash
Touch contaminated tray
Spoon and right gripper: contaminated
Real-robot case
Household robot planning, with a memory of contact
HygieneRoboBench
Benchmarking Hygiene-Aware Planning for Household Robots
Real-robot case + native simulation 2×
A world of household decisions.
TL;DR
A robot’s next safe action depends on what it touched before. HygieneRoboBench tests whether planners can use contact history to choose safe, cost-optimal continuations under user priorities and newly revealed events.
624task instances
134task families
5activity groups
7household areas
01 The problem
Same scene. Different next move.
A washed gripper receives a spoon. Can it handle fruit next? The answer depends on the contacts that came before.
A contact changes the continuation.
The formal task compares washing and contact-pad replacement under the user’s priorities.
Explore the same decision pointFixed task, resources and spare-parts-first priority
History 2 Wash, then touch
Touch contaminated tray
Spoon and right gripper: clean
Put down → Wash → Place fruit
Safe and optimal
Put down → Place fruit
Safe and optimal
State colors illustrate benchmark rules.
Actions and schedules stay fixed. Only the history used for evaluation changes.
02 The benchmark
Everyday tasks. Hidden consequences.
Five activity groups adapted from BEHAVIOR-1K, with explicit contact histories, treatment requirements, and user priorities.
Across seven areas
Native OmniGibson illustrations of benchmark tasks.
One dataset. Three ways to understand it.
Each ring partitions the same 624 instances.
- Outer ringHousehold activities
- Middle ringConstruction designs
- Inner ringPlanning difficulty
What should change the plan?
Contact history
Change designated contact records while keeping priorities and task conditions fixed.
User priorities
Keep action costs fixed. Change which resource the user wants to conserve first.
New events
Reveal a new contact, then evaluate the continuation from that decision point.
Shared definitions
One task. Five connected registries.
Explore the definitions behind the fruit task, from its service goal to the costs of treatment.
From shared definitions to controlled tasks.
Vary history, priorities, or event information within a task family, then check the resulting instances.
Task sources
A clear service goal
Place fruit on a plate.
01Task sources
Adapt the activity while preserving the service goal and functional object roles.
02Hygiene rules
Specify contact transfer and treatment effects. Materials and procedures determine the applicable treatment conditions.
03Events
Represent newly revealed contacts or changes. Observed contacts enter the history used to reconstruct the current state.
04Actions & resources
Define targets, preconditions, durations, and resource use. The wash requires an empty gripper and treats only its contact part.
05Costs & priorities
Compare feasible plans using the user’s ordered cost priorities. Saving spare parts and saving water can favor different actions.
From definitions to decisions
The same costs. A different priority.
Washing saves spare inserts. Replacement saves water. All planners use the same abstract task parameters.
Preferred treatment: Wash
| Action | Time, TU | Water | Cleaner | Spare parts | Solid waste |
|---|---|---|---|---|---|
| Wash | 3 | 2 | 1 | 0 | 0 |
| Replace | 2 | 0 | 0 | 1 | 1 |
Sources, modeling choices, and human review +
External basis & formalization
Hygiene studies and public-health guidance inform material and procedure requirements. Authors formalize these principles as discrete transfer and treatment rules.
Parameters & review
The general construction process screens online time and supply estimates with an LLM, then checks plausibility and relative magnitudes through human review. The fruit values above are adopted abstract modeling parameters.
Two reviewers check task wording, applicability, action conditions, and consistency; a third provides final review. Program checks validate matched conditions, feasibility, and reference costs.
Evaluation
A common test for every planner.
The evaluator replays history, checks the submitted continuation, and compares costs against a private reference under the same priorities.
Public input
→Request + context
History, resources, events, priorities, shared rules
Planner output
→Continuation
Actions and schedule, or an infeasibility judgment
Private assessment
Replay + reference
Constraints, recomputed costs, and reference targets
Washing requires an empty gripper.
SR
Safe resolution
Safe completion, or correct rejection of a certified-infeasible continuation.
OSR
Optimal safe resolution
SR with reference-optimal cost for a submitted plan, under the user’s priorities.
03 The findings
Safety is one question. Efficiency is another.
Seven planners, one evaluation protocol. Examine overall performance, task difficulty, and activity-specific strengths.
Overall resolution, 624 instances
0%50%100%
Hygiene-NSP achieves the highest overall SR and OSR in this evaluation.
98
Safe, nonoptimal DeepSeek plans. In 41 cases, water or spare parts is the first nonoptimal cost dimension.
7 / 7
Planners score lower as difficulty increases. Activity breakdowns reveal different strengths; Gemini leads cleaning and laundry.
Controlled history swaps
Change what matters. Control what does not.
Swap histories while preserving saved actions and schedules. Relevant swaps test history-dependent decisions; irrelevant swaps provide the control.
A separate task: move the bowl to the sink.
At the decision point: clean grippers, different side-pad states.
Control history 1
Wash gripper → Touch pad
Control history 2
Touch pad → Wash gripper
Different pad states. Same continuation.
Neither saved plan touches the pad again.
Safety and optimality unchanged
Relevant history0 / 1,699
Originally safe, optimal replays retain both after swapping.
Irrelevant history1,288 / 1,288
Safety and optimality outcomes remain unchanged.
All seven planners, including 15 newly safe, optimal outputs +
User priorities
234 / 268 general pairs
24 / 25 ranking-flip pairs
Hygiene-NSP achieves the highest pair-success counts in both groups. Action costs stay fixed while priorities change.
Contact events
50 relevant + 50 irrelevant
Hygiene-NSP reaches 88% and 92% OSR. Gemini reaches 90% and 94%, the highest in both event conditions.
All contact-event results +
04 The companion planner
Hygiene-NSP
Ground the request, reconstruct the contact-dependent state, and jointly plan treatment and service actions under user priorities.
01
Language grounding
Goals, objects, and priorities
02
Trusted compilation
Definitions and constraints
03
History replay
Current hygiene state
04
CP-SAT planning
Treatments and schedules
Optional state projection
A smaller planning representation.
Retain task-relevant information, with full-state fallback when necessary.
20.1%
Median family-level reduction in local planning time, excluding LLM calls.
175 / 177
Paired runs agree on status and cost, including 21 full-state fallbacks.
The complete study
Watch the research film.
From the real-robot example to the benchmark, controlled evidence, and companion planner.
Read the narrated transcript +
Explore the work
Paper & resources.
Read the paper, watch the research introduction, and explore the repository.