01→
Observe
Combine visual, spatial, temporal, environmental and human-provided signals.
Sinuosity Physical Intelligence Lab
Building intelligence for the physical world.
Sinuosity develops multimodal models and autonomous systems that perceive, reason and operate in complex real-world environments. We begin in agriculture, where variability, constrained connectivity and consequential decisions demand a more rigorous form of intelligence.
INPUTSVisual · spatial · temporal
MODEL BEHAVIORObserve · calibrate · request evidence
Multimodal LearningEmbodied IntelligenceEdge AIField RoboticsLanguage InterfacesSafety & Evaluation
Our research thesis
The physical world remains an unsolved intelligence problem.
AI systems have become increasingly capable in language, code and other digital domains. Physical environments remain harder. Conditions change continuously. Sensors provide incomplete information. Hardware imposes strict constraints. Actions can damage equipment, crops or surrounding environments.
We research the models and systems required for AI to understand these conditions, act under uncertainty and improve through real-world experience.
02→
Understand
Represent objects, conditions, relationships, changes and uncertainty.
03→
Plan
Choose the next observation or action under operational and safety constraints.
04→
Act
Connect high-level intelligence to deterministic control systems and human oversight.
05→
Learn
Improve from outcomes, corrections, interventions and repeated field experience.
Research
Research programs
Field Foundation Models
Multimodal models that represent crops, terrain, field conditions and change over time using ground imagery, video, geospatial information, weather, metadata and human observations.
- Agricultural visual representation learning
- Multi-image and video understanding
- Temporal field reasoning
- Crop and terrain segmentation
Embodied Intelligence
Learning systems that connect perception and reasoning to physical action, beginning with agricultural machines operating in unstructured outdoor environments.
- Vision-language-action models
- Teleoperation-to-autonomy
- Learning from human intervention
- Outdoor navigation
Edge and Offline AI
Efficient models and runtimes designed to operate locally where bandwidth, latency, power, privacy or reliability make continuous cloud access unsuitable.
- Model compression
- Quantization
- On-device multimodal inference
- Hybrid edge-cloud systems
Language and Human Interfaces
Speech and language systems that support natural interaction with scientific and physical-intelligence systems, beginning with Twi, Ghanaian English and real-world code-switching.
- Low-resource speech recognition
- Natural text-to-speech
- Tonal-language modeling
- Code-switching
Safety and Real-World Evaluation
Evaluation methods for determining whether intelligent systems remain reliable, calibrated and safe outside controlled demonstrations.
- Field generalization
- Human intervention measurement
- Failure and recovery analysis
- Physical safety boundaries
Applied intelligence / Introducing FAMA
Crop intelligence.
In hand.
Meet FAMA, Okuafo’s compact handheld for crop intelligence. A rear camera, a speaker and three tactile controls bring an offline-first, audio-first approach closer to the farmer.
Models and systems
Research translated into working systems.
Sinuosity develops research models, evaluation frameworks and experimental platforms together. Models are studied through data-collection and field systems rather than evaluated only on static laboratory datasets.
Every record below states its current lifecycle and evidence basis. No item is presented as an open release unless access is explicitly published.
Okuafo MaizeGuard Edge v1.5
Calibrated two-model maize symptom screening system
A public development candidate combining categorical and specialist FP16 TensorFlow Lite models with locked fusion calibration and explicit abstention for offline, human-reviewed maize-leaf symptom screening.
Okuafo MaizeGuard Edge v1.4
Two-stage offline maize screening system
A research-preview, two-stage LiteRT and TensorFlow Lite vision system evaluated on documented development datasets and limited field samples for offline, human-reviewed maize-leaf screening.
FieldMind-1
Multimodal field foundation model
FieldMind-1 is a family of multimodal models being developed to understand agricultural environments across images, video, metadata, location and time.
FieldAction-1
Embodied agricultural policy model
FieldAction-1 investigates how agricultural machines can learn useful field behaviors from teleoperation, demonstrations, interventions, failures and experience across platforms.
TwiField Speech
Low-resource speech and voice model family
TwiField Speech investigates speech recognition and generation for Asante Twi, Akuapem Twi, Ghanaian English and agricultural code-switching.
FieldBench
Real-world physical-intelligence evaluation framework
FieldBench is being designed to evaluate autonomous and multimodal systems under real agricultural operating conditions.
Farmba Scout
Experimental field-intelligence platform
Farmba Scout is a modular agricultural rover prototype used to study field data collection, perception, teleoperation-to-autonomy and offline physical-intelligence workflows.
Technical system
From field observation to physical intelligence.
Reference research architecture · illustrative
01
Data acquisition
- Smartphones
- Field agents
- Rover cameras
- Depth sensors
- Drones
- Weather
- Soil
- Geospatial data
- Voice
02
Data engine
- Provenance
- Quality control
- Deduplication
- Pre-labeling
- Expert review
- Temporal linking
- Outcomes
03
Models
- Okuafo MaizeGuard Edge
- FieldMind
- FieldAction
- TwiField Speech
- Uncertainty models
04
Deployment
- Mobile
- Cloud API
- Edge computer
- Rover
- Farm dashboard
- Partner machinery
05
Evaluation
- FieldBench
- Private test sets
- Safety tests
- Intervention tracking
- Longitudinal outcomes
Initial proving ground
Agriculture is our first proving ground.
Agricultural environments expose many of the central problems of physical intelligence at once. Biological systems change over time. Symptoms may have multiple causes. Fields differ by soil, crop, weather, management practices and camera conditions. Connectivity is not guaranteed, and incorrect decisions can carry real financial and environmental costs.
A system that performs reliably under these conditions must do more than recognize an image. It must integrate context, represent uncertainty, request better evidence, operate efficiently and learn from what happens after a decision is made.
Our agricultural work is both an applied mission and a rigorous research environment for developing intelligence that can eventually extend to other outdoor and industrial domains.
Data and learning
Models improve through structured experience.
The central research asset is not a collection of disconnected photographs. It is a longitudinal record connecting observation, interpretation, action and outcome.
Observation
Field imagery, sensor readings and human reports.
Interpretation
Visible evidence, candidate explanations and uncertainty.
Action
Inspection, treatment, human intervention or machine behavior.
Outcome
Subsequent field condition, recovery, failure, yield and cost.
Data collection must be consented, provenance-aware and governed according to clear internal, research and public-release permissions.
Read our data governance approachOpen research
Open where it advances the field.
Sinuosity releases selected model artifacts, evaluation tools, reference implementations, technical reports and carefully governed datasets where public access can accelerate progress without compromising privacy, rights, safety or evaluation integrity.
Dataset release / 002
Ghana Agricultural Data Commons v0.2.0
Ten public Hugging Face dataset repositories for Ghana agricultural data, with registry, provenance, governance, redistributable crop-statistics payloads where allowed and GhanaAgBench benchmark assets.
Public access does not imply a blanket open-source license grant. Review each repository license, source receipt and limitation before reuse.
Safety
Physical intelligence requires layered safety.
Models operating in physical environments must be evaluated differently from ordinary software. Sinuosity studies layered safeguards spanning model behavior, deterministic control, hardware limits, human oversight and operational procedures.
Semantic safety
Understanding instructions, restrictions and unsafe requests.
Decision safety
Uncertainty calibration, abstention and human approval.
Control safety
Deterministic motion limits, geofencing and emergency stopping.
Operational safety
Training, maintenance, permissions and incident procedures.
Evaluation safety
Independent testing, private benchmarks and failure reporting.
Research register
Research and technical updates.
Technical ReportResearch NoteResearch PreviewDatasetField StudyEngineeringSafetyAnnouncement
Model Release / 2026
Field vision / calibrated screening
Okuafo MaizeGuard Edge v1.5
A public development candidate for human-reviewed maize symptom screening, combining categorical and specialist FP16 TFLite models through a locked calibrator with explicit abstention.
- System
- Categorical + specialist fusion
- Runtime
- FP16 TensorFlow Lite
- Decision
- Calibrated abstention
- Status
- Public development candidate
Published for user testing with a known gray leaf spot recall limitation. It is not independently or agronomist validated and is not approved for production or autonomous diagnosis.
Collaboration
Work with Sinuosity.
We collaborate with researchers, universities, farms, equipment manufacturers, scientific institutions and technology partners working on difficult physical-world intelligence problems.
Research institutions
Joint studies, graduate research, benchmarks and publications.
Farms and agricultural organizations
Field sites, longitudinal data and operational evaluation.
Robotics and equipment companies
Hardware integration, autonomy and cross-platform learning.
Compute and technology partners
Cloud infrastructure, edge hardware, sensors and developer tools.
Scientific specialists
Agronomy, plant pathology, soil science, remote sensing and linguistics.