01
Predictive Physical AI
Learning compact representations of motion, dynamics, and change from video — moving beyond frame-by-frame recognition toward predictive understanding.
PhyNexa Labs · Physical Intelligence
Physical AI,
Everywhere.
We develop efficient AI that perceives, predicts, and understands the physical world — bringing advanced intelligence closer to every device, system, and application.
PERCEPTION/PREDICTION/EFFICIENCY/PHYSICAL INTELLIGENCE
01 / OUR TECHNOLOGY
Intelligence that meets the real world.
Three connected research directions: predictive learning, grounded perception, and compute-efficient intelligence.
02
Grounded Multimodal Perception
Connecting vision and language with real-world evidence for more reliable visual reasoning, understanding, and industrial inspection.
03
Efficient AI Systems
Advancing model and inference efficiency so powerful physical intelligence can reach smaller GPUs, CPUs, and on-device environments.
Multimodal Grounding & Hallucination Reduction
Vision-language models can generate plausible descriptions that are not supported by visual evidence. We investigate grounding-aware supervised fine-tuning and preference optimization to improve factual reliability and reduce hallucinations. Our research focuses on aligning model responses with observable image content, supported by curated training data, evaluation benchmarks, and experimental results.
Click the slide to enlarge.
OCR-Aware Visual Understanding
Reading text in real-world images is challenging due to blur, occlusion, perspective distortion, and multilingual content. We develop OCR-aware multimodal post-training methods to improve text extraction and grounded visual reasoning. In our experiments, information completeness improved from 71.3 to 84.6, while hallucination decreased from 18.3% to 5.5%, using the same frozen visual encoder.
Click the slide to enlarge.
Sensor Fusion & Spatiotemporal Prediction
Physical-world measurements are often sparse, noisy, and distributed across multiple sensors, locations, and time periods. We investigate sensor fusion, spatial statistical modeling, and temporal data analysis to reconstruct and predict physical properties. Prior applications include intelligent road construction, where heterogeneous measurements support spatial prediction and data-driven understanding of material and infrastructure conditions.
Click the slide to enlarge.
PRODUCTS
What we are building.
Compact VLM Classification & Reasoning
Large vision-language models can be computationally expensive to deploy, particularly for specialized visual understanding tasks that require both accurate classification and interpretable reasoning. We develop compact VLM architectures with task-specific classification and reasoning capabilities, combining efficient multimodal representations with specialized post-training strategies. In our evaluation, an 8B model achieved an F1 score of 0.863, outperforming a larger 31B model (0.834) on 1,500 held-out expert-labeled examples.
Click the slide to enlarge.
Lightweight Adaptive Anomaly Detection
Industrial defect detection requires reliable inspection with minimal computational overhead. We combine normalized cross-correlation (NCC) template matching with statistical KDE and image-adaptive fusion to provide a lightweight alternative to deep neural networks. On ceramic debris detection, our method achieved 95.2% F1, compared with 99.3% for ResNet-50, demonstrating competitive accuracy without neural-network training.
Click the slide to enlarge.
Automated Model Evaluation & Checkpoint Selection
Selecting reliable model checkpoints from extensive post-training experiments is costly and difficult when evaluation depends on inconsistent or expensive manual judgments. We develop an automated evaluation framework that progressively combines pointwise scoring, listwise ranking, and pairwise comparison to assess model quality and identify promising checkpoints. The framework supports systematic, repeatable model selection and reduces the need for exhaustive manual comparisons.
Click the slide to enlarge.
Newton-Based Optimization
Hyperparameter optimization and post-training can require many expensive evaluations, slowing model development and increasing computational costs. We explore Newton-inspired numerical optimization methods that use local objective information to guide efficient parameter updates and search. In a reported XGBoost experiment, the approach required 8 optimization iterations versus 320 grid-search evaluations to reach competitive performance. The framework also targets efficient optimization in VLM post-training and classical machine learning.
Click the slide to enlarge.
02 / APPLICATIONS
Built for practical impact.
From industrial visual intelligence to efficient on-device perception, we are developing AI that can operate where the physical world happens.
01
Industrial visual inspection
Perception and anomaly detection for real-world quality and reliability challenges.
02
Predictive video understanding
Learning useful temporal representations from visual experience.
03
Accessible multimodal AI
Efficient vision-language intelligence for constrained compute and edge environments.
03 / RESEARCH
Research-driven innovation.
Selected foundational research related to PhyNexa's technical directions. Publications are attributed to their original authors and institutions; they are not presented as company-owned intellectual property.
Predictive representation learning
Momentum-Guided Semantic Forecasting (MoFore)
Self-supervised video representation learning through forecasting in latent space.
Grounded multimodal AI
Reducing Hallucination in Multimodal Large Language Models through Hard Grounding Preference Supervision
Research on improving grounded and reliable vision-language reasoning.
Efficient optimization
Efficient Constrained Graph Search for Post-hoc Error Correction in Binary Classifiers
Algorithmic approaches to efficient and accurate decision systems.
04 / CONNECT
Let's build what's next.
Interested in product service, research collaboration, technology partnerships, or learning more about PhyNexa Labs? Reach us at qxu@phynexa.net.