Vision-language-action models
Fine-tuning and adapting VLA models so robots can turn camera input and instructions into actions.
- VLA
- imitation learning
- fine-tuning
- household tasks
Torquero LabAI software & models for robotics
AI that
AI that
moves.
Torquero Lab is our applied research and development lab for AI software and LLM-based models for the robot industry. We help robots understand language, see the world, plan tasks and act on them, on real hardware.
# from words to motion, on the robot robot: unitree-g1 perceive: stereo-rgb + depth + lidar reason: llm-planner # language → task graph act: vla-policy # vision-language-action control: rl-locomotion # trained in sim deploy: jetson-orin # quantized, on-device eval: sim → real # safety checks first ✓ "pick up the red block and bring it to me"
Focus areas
The software layer that makes robots useful.
Hardware is getting better fast. The difference between a demo and a robot that does real work is the software: models that understand, plan and act safely in the physical world.
Language, reasoning & task planning
LLM-based agents that break natural-language requests into robot skills, with tool use and guardrails.
- LLM planning
- tool calling
- voice interaction
Locomotion & whole-body control
Reinforcement-learning policies trained in simulation and transferred to humanoids and quadrupeds.
- RL
- sim-to-real
- whole-body control
Perception & world models
Scene understanding from cameras, depth and LiDAR: detection, mapping and models of how the world changes.
- 3D perception
- SLAM
- world models
Robot data & teleoperation
Pipelines to collect, clean and label robot data, from teleoperation sessions to simulation-generated datasets.
- teleoperation
- synthetic data
- datasets
On-device AI & evaluation
Compressing and running models on the robot's own compute, with test suites that check behaviour before it ships.
- edge inference
- quantization
- benchmarks
How we build
From model to motion.
Every project follows the same loop, so models move from experiment to real robot in a controlled way.
01
Data
Collect demonstrations, teleoperation and sensor logs.
02
Train
Fine-tune foundation models and train control policies.
03
Simulate
Test at scale in physics simulation before touching hardware.
04
Deploy
Optimise for the robot's onboard compute and integrate with its SDK.
05
Evaluate
Measure on real tasks, check safety, and feed results back into data.
Work with the lab
Built for the people building robots.
Custom AI development
Models, skills and software stacks for your platform, from language interfaces to new behaviours.
Research partnerships
Joint projects on embodied AI, with hardware from our catalog and a team that knows the platforms.
Pilots & proofs of concept
Scoped projects that test whether a robot and AI approach works for your use case before you scale.
Training & workshops
Hands-on sessions on robot AI, LLM agents and sim-to-real, built around the robots in our catalog.
Principles
How we work.
Safety firstSimulation and evaluation before real-world tests, with emergency stops always in the loop.
Hardware-agnosticSoftware that works across makers, not locked to a single robot.
On-device firstRun models on the robot where possible, for speed, reliability and privacy.
Open where it helpsBuild on open models, tools and standards, so customers keep control of their stack.
ROS 2NVIDIA Isaac SimMuJoCoPyTorchJetson Orin / ThorPython & C++ SDKsURDF / MJCF
FAQ
Questions about the lab.
What does Torquero Lab do?
Torquero Lab is Torquero's applied research and development lab for AI software and LLM-based models for the robot industry. It works on vision-language-action models, LLM agents and task planning, locomotion and sim-to-real transfer, perception, robot data pipelines and on-device deployment.
What is sim-to-real?
Sim-to-real means training a robot's behaviour, such as walking or grasping, in a physics simulator and then transferring it to the real robot. Simulation lets a robot practise far more than it could in the real world, safely and quickly.
Can LLMs and AI models run on the robot itself?
Often, yes. With compression techniques such as quantization, many models can run on a robot's onboard computer, for example NVIDIA Jetson Orin or Thor modules. Running on-device improves speed, reliability and privacy.
Which robots does Torquero Lab work with?
The lab is hardware-agnostic. It works across makers and with the humanoids, robot dogs, hands and actuators in the Torquero catalog, including platforms from Unitree, Booster Robotics, AgiBot and RoboParty.
How do we start a project with Torquero Lab?
Email lab@torquero.com with your robot platform and what you want it to do. Projects range from research partnerships and custom AI development to pilots, proofs of concept and training workshops.
Have a robot that should be smarter?
Tell us about your platform and what you want it to do. We will tell you how the lab can help.