Capabilities
Six modules, from raw sensor data to action.
Each module solves one part of a Physical AI system. Use the complete stack, or engage us for the modules your product needs.
01 · Capture
Data
Physical AI learns from the physical world. This module records how people actually do the task and what the product's sensors pick up, then labels and structures it for training.
Includes
- Human demonstration capture
- Wearable and sensor recording
- Labelling and dataset structuring
- Field data fed back after launch
Example
Record operators packing boxes. Get a dataset a robot arm can learn from.
02 · Sense
Perception
Raw frames, audio and sensor readings mean nothing on their own. This module turns them into events: a person arrived, a word was spoken, an object was touched.
Includes
- Speech recognition and wake words
- Person and object detection
- Touch, motion and IMU events
- Fusion across sensors
Example
A companion robot hears its name, sees a person and feels a touch, as one event.
03 · Think
Cognition
This is where the product decides. It remembers past interactions, knows the current context, works toward a goal and checks every action against limits you set. It runs on Recapi, our cognitive layer.
Includes
- Short and long-term memory
- Context and domain knowledge
- Goal planning and decisions
- Guardrails on speech and action
Example
A reception robot remembers a returning visitor and the room they need.
04 · Learn
Models
We train and fine-tune small language models and multimodal models for the task, test them for accuracy and behaviour, and compress them to run on the hardware inside your product.
Includes
- SLM and multimodal model selection
- Fine-tuning on domain data
- Accuracy, safety and edge-case tests
- Optimisation for on-device inference
Example
A speech model that runs on the robot's own chip, with no cloud round trip.
05 · Act
Control
A decision only matters once the hardware carries it out. This module connects the cognition layer to speakers, motors and actuators, and handles timing and safety limits.
Includes
- Voice and speech output
- Motor and actuator interfaces
- Timing from sensing to action
- Safety limits on movement
Example
Decide to hand over a part, then move the arm at a safe speed to do it.
06 · Own
Your AI
Physical products can't always rely on the cloud. Your AI keeps the intelligence with the machine: a compact model on the robot or device, or on your own servers, that stays linked to the wider Recapi network for updates. We monitor it, secure it and ship controlled updates.
Includes
- Compact models on the device
- Private, on-site hosting
- Linked to the wider Recapi network
- Monitoring and controlled updates
Example
A factory system that keeps working when the internet drops.
Across every layer
Evaluation runs across every layer, measuring perception, model performance, decision quality and physical execution.
Around the modules
Compute and hardware sourcing through our partners, and forward-deployed engineers inside your team.
Let's build together
Have a physical product in mind?
Let's make it intelligent.
You bring the product idea. We bring the intelligence, data and engineering to make it work.