04 / RSI / IOT TASK BENCH
RSI-IoT algorithm research tasks
RSI Task Bench turns real perception data into trainable and evaluable environments for algorithm-research agents.
The current scope is RSI-IoT. IMU time series, images, radar, depth data, and speech features are organized into tasks with explicit data boundaries, resource limits, and acceptance checks. Within a fixed compute and time budget, an agent analyzes data, designs a method, validates it experimentally, and submits a runnable solution rather than a written proposal alone.
RSI TASK BENCH / IOT
Perception tasks in scope
Six task forms spanning time series, vision, radar, depth, and speech.
- 01
IMU activity recognition
Build a reproducible classifier from time-series windows and activity labels.
- 02
Cross-client generalization
Hold out clients to test transfer across devices and participants.
- 03
Multimodal few-shot learning
Combine limited labeled examples from several perception modalities.
- 04
Infant-monitor image classification
Classify authorized monitor images under a defined label set.
- 05
Radar and image person detection
Design a detector and validate predictions against held-out scenes.
- 06
5-shot spoken commands
Adapt a speech-command recognizer from five examples per class.
DELIVERY / CONTENT
What a delivery can include
Defined with each project and its authorization scope.
- 01Harbor isolated environment, public development data, and a separate hidden test set
- 02Task brief, unified inference entry point, experiment logs, and runnable solution
- 03Public scorer and independent automated evaluator, with task-appropriate Macro F1, mean cross-client accuracy, or AP50 feedback
SCENARIO / EXAMPLE
A task in context
For cross-client IMU activity recognition, an agent examines class and client differences in public development data, trains a model, and submits through a unified inference entry point. Held-out clients remain reserved for independent evaluation.
What we verify
Check task-appropriate Macro F1, mean cross-client accuracy, or AP50, plus time and compute limits, data separation, executable inference, and reproducibility.