cs.AIApr 15, 2026
Save[Emerging Ideas] Artificial Tripartite Intelligence: A Bio-Inspired, Sensor-First Architecture for Physical AI
Organizations: Seoul National University, Seoul, South Korea
Abstract
As AI moves from data centers to robots and wearables, scaling ever-larger models becomes insufficient. Physical AI operates under tight latency, energy, privacy, and reliability constraints, and its performance depends not only on model capacity but also on how signals are acquired through controllable sensors in dynamic environments. We present Artificial Tripartite Intelligence (ATI), a bio-inspired, sensor-first architectural contract for physical AI. ATI is tripartite at the systems level: a Brainstem (L1) provides reflexive safety and signal-integrity control, a Cerebellum (L2) performs continuous sensor calibration, and a Cerebral Inference Subsystem spanning L3/L4 supports routine skill selection and execution, coordination, and deep reasoning. This modular organization allows sensor control, adaptive sensing, edge-cloud execution, and foundation model reasoning to co-evolve within one closed-loop architecture, while keeping time-critical sensing and control on device and invoking higher-level inference only when needed. We instantiate ATI in a mobile camera prototype under dynamic lighting and motion. In our routed evaluation (L3-L4 split inference), compared to the default auto-exposure setting, ATI (L1/L2 adaptive sensing) improves end-to-end accuracy from 53.8% to 88% while reducing remote L4 invocations by 43.3%. These results show the value of co-designing sensing and inference for embodied AI.
Figures & tables
Figure 1. Limitations of prevailing paradigms for physical AI. (a) Computation-centric approach: sensors are treated as passive inputs while adaptation is confined to the model stack. (b) Cortex-centric adaptive sensing: sensing is controlled directly by a monolithic model without structured reflex and calibration layers. Two-panel schematic comparing physical AI paradigms: computation-centric versus cortex-centric adaptive sensing.
| Modality | Human sensor (adaptive) | Artificial sensor (static) |
| Vision | Pupil dynamically controls light intake. Photoreceptors adapt to luminance over seconds. Fast saccades redirect the fovea before cortical processing. | Camera (controls: exposure, ISO, frame rate, focus). Frame rate and exposure/ISO set in coarse steps, rarely changed. Saturation and motion blur corrected mainly after capture. |
| Hearing | Active gain control via outer hair cells over a wide dynamic range. Efferent feedback protects against loud sounds. Sub-millisecond binaural timing enables precise localization. | MEMS microphone (controls: gain/AGC, sampling rate, beam pattern). AGC, sampling rate, and bit depth fixed at design time. 3D localization relies on multi-mic arrays and heavy DSP. |
| Touch | Dense, compliant skin with high spatial and temporal acuity. Reflex arcs withdraw from painful stimuli before awareness. Active exploration (rubbing) increases signal-to-noise ratio. | Tactile array / pressure sensor (controls: readout freq, active ROI). Sparse capacitive or piezoresistive taxel arrays on rigid surfaces. Thresholds and sampling rates fixed; protection left to high-level control. |
| Proprioception | Vestibular and proprioceptive organs track body pose and acceleration. Rapid reflexes stabilize posture and gait without cortical involvement. | IMU (controls: measurement range, sampling rate, filter parameters). Accel/gyro data sampled at fixed rates with offline bias calibration. Ranges and filters rarely change after deployment. |
Table 1. Adaptive human sensing versus conventional artificial sensors across four modalities.
| Biological inspiration | Functional role in biology | ATI component |
| Brainstem | Reflex and safety control | L1 safety reflex and signal protection |
| Cerebellum | Predictive calibration and error correction | L2 sensing calibration and sensor control |
| Basal Ganglia Network (BGN) | Skill selection and habitual execution | L3 task-specific skill selection/execution on device |
| Frontoparietal Network (FPN) | Flexible coordination across systems | Routing (L3–L4 coordination) |
| Hippocampal–Cortical Network (HCN) | Memory and relational reasoning | L4 deep reasoning and knowledge-based inference |
Table 2. Biological inspirations as conceptual motifs for ATI roles.
Figure 2. ATI architecture. (a) ATI role mapping for a camera-based system. (b) General ATI interfaces, from physical-world sensor inputs to on-device processing and edge/cloud reasoning. ATI architecture overview: role mapping for a camera-based system and general interfaces from sensing to on-device and edge or cloud reasoning.
| Function | Description |
| set_shutter(step) | Exposure duration (discrete shutter step) |
| set_gain(dB) | Analog signal amplification (in dB) |
| set_hdr(mode) | HDR / WDR bracketing strategy |
| set_roi(box) | Active pixel region read-out (ROI) |
| slew_gimbal( p, t) | Pan/Tilt actuation velocity |
Table 3. Example sensor control APIs exposed by ATI for camera and gimbal control.
Figure 3. Experimental setup for the prototype. A Galaxy S25 is mounted on a small vehicle that drives along a closed track. Ambient illumination is set to bright (left) and dark (right). Prototype setup with a smartphone-mounted vehicle on a closed track under bright and dark lighting conditions.
Figure 4. L2 sensor calibration via contextual bandits: learning and consolidated policy lookup.
| Category | Control Method | Local-Only | Remote-Only | Split/Routing |
| Baseline | Default AE | AE-L3 | AE-L4 | AE-L3-L4 |
| Baseline | Vendor EIS | EIS-L3 | EIS-L4 | EIS-L3-L4 |
| ATI | L1 (Reflex Only) | L1-L3 | L1-L4 | L1-L3-L4 |
| ATI | L1/L2 (Full Stack) | L1/L2-L3 | L1/L2-L4 | L1/L2-L3-L4 |
Table 4. Evaluated configurations formed by combining sensing strategies and inference paths.
Figure 5. Evolution of camera parameters and confidence scores during L2 learning under low illumination. L2 adapts ISO and exposure within the safety envelope of L1. The rolling average of confidence score (red line) becomes higher and more stable over time, indicating input stabilization. Time-series plot showing ISO, exposure, and confidence during L2 learning under low illumination, with increasing confidence over time.
| Method | Model Config | Per-Class Accuracy (%) | Total | L4 Call | |||||||
| Teddy | Racket | Tennis Ball | Ping-pong Ball | Orange | Carton | Water Bottle | Laptop | Rate (%) | |||
| Baselines | AE-L3 | 84.0 | 32.0 | 90.0 | 94.0 | 16.0 | 28.0 | 34.0 | 22.0 | 50.0 | - |
| AE-L4 | 94.0 | 26.0 | 86.0 | 18.0 | 76.0 | 84.0 | 86.0 | 80.0 | 68.8 | 100.0 | |
| AE-L3-L4 | 84.0 | 32.0 | 88.0 | 88.0 | 22.0 | 38.0 | 44.0 | 34.0 | 53.8 | 56.0 | |
| EIS-L3 | 46.0 | 0.0 | 100.0 | 0.0 | 0.0 | 2.0 | 8.0 | 0.0 | 20.0 | - | |
| EIS-L4 | 94.0 | 4.0 | 86.0 | 0.0 | 76.0 | 10.0 | 88.0 | 46.0 | 51.0 | 100.0 | |
Table 5. Classification accuracy and L4 usage across twelve configurations formed by combining four sensing strategies (AE, EIS, L1, L1/L2) with three inference paths (L3, L4, and L3-L4). Per-class and total accuracies are reported over eight object classes. L3 uses EfficientNet-Lite0 as the local model, and L4 denotes the remote inference path evaluated via offline replay. The L4 call rate is measured on a per-lap basis and reports the percentage of laps processed by L4.
Figure 6. Sample images of the eight object classes used in the experiment. Sample images of the eight object classes used in the experiment.
| Method | Model Config | MobileNet V2 | MobileNet V3 |
| Baselines | AE-L3 | 42.8 | 28.5 |
| AE-L3-L4 | 48.5 (L4 call 58.8%) | 34.8 (L4 call 52.3%) | |
| ATI variants | L1-L3 | 58.8 | 51.0 |
| L1-L3-L4 | 62.3 (L4 call 33.8%) | 52 (L4 call 37.8%) | |
| L1/L2-L3 | 66.0 | 53.3 | |
| L1/L2-L3-L4 (Ours) | 71.5 (L4 call 29.0%) | 64.8 (L4 call 47.8%) |
Table 6. Classification accuracy across baseline and ATI configurations with different local backbones. L4 call rates for hybrid L3-L4 settings are shown in parentheses.
Figure 7. Ablation of the L3 escalation threshold (EfficientNet-Lite0): (Left) Accuracy. (Middle) Escalation rate. (Right) Accuracy–escalation trade-off; best at . Three plots showing accuracy, escalation rate, and their trade-off across L3 escalation thresholds, with the best result at .
Figure 8. Performance under dynamic lighting on the Teddy object. (a) Time-series traces of illuminance, exposure time, ISO, and sharpness for Auto Exposure (AE, blue) and ATI (black) under alternating bright and dark conditions. (b) Comparison of L3 accuracy (left) and L4 call rate/final accuracy (right) between AE and ATI. Comparison of Auto Exposure and ATI under dynamic lighting: time-series traces on the left and cascade performance metrics on the right.
Figure 9. Qualitative comparison of captured frames under different sensing strategies: Auto Exposure (AE), vendor electronic image stabilization (EIS), L1, and L1/L2. Qualitative comparison of captured frames under four sensing strategies: Auto Exposure, vendor electronic image stabilization, L1, and L1/L2.
Figure 10. Auditory ATI proof of concept. (a) Physical audio setup with one microphone, one signal speaker, and one noise speaker. (b) SNR distributions for AGC and ATI under a noise-dominant condition. Auditory ATI proof of concept: physical audio setup on the left and SNR distributions for AGC and ATI on the right.