Research

Applied research in perception, learning, simulation and manipulation, aimed at robots that work reliably outside controlled demonstrations.

The themes on this page are Axinome’s research focus areas: the questions we are organised to investigate and the methods we use. They are not a record of completed results.

Robot perception and autonomy

Scope: Sensor fusion, environment understanding, localisation, mapping, navigation and task execution.

Research questionHow can a robot keep an accurate, usable model of its surroundings when sensors are noisy, conditions change and parts of the scene are unknown?

Why it matters: Most failures in deployed mobile robots begin with a wrong estimate of position or of the scene. Reliable perception is what allows a robot to move and act without constant supervision.
How we approach it: We combine camera, depth, LiDAR and inertial data, compare estimation methods against ground truth where it can be measured, and study how errors propagate into planning and task execution.
What progress looks like: Lower localisation drift, fewer navigation failures in changing scenes, and behaviour that degrades predictably when a sensor is lost.

Simulation render of an Intel RealSense D435i depth camera seen from the front, showing its lenses. — Simulation render Intel RealSense D435i depth camera, an example of the sensors we integrate. Simulation render by Axinome.

Robot learning and physical AI

Scope: Learning from demonstrations, reinforcement learning, embodied decision-making and the evaluation of learned behaviours.

Research questionWhen does a learned behaviour work well enough to rely on in a physical task, and how do we show that with evidence rather than a single successful demonstration?

Why it matters: Learning methods can produce motions and decisions that are hard to program by hand. They also fail in ways that are hard to predict, so evaluation matters as much as training.
How we approach it: We treat learned components as parts of a larger system: trained against defined tasks, compared with conventional baselines, and tested across repeated trials and varied conditions.
What progress looks like: Learned behaviours with measured success rates, known failure modes and clear conditions under which they should not be used.

Simulation and transfer to hardware

Scope: Simulation environments, synthetic data, scenario testing and assessing how simulated behaviour transfers to physical robots.

Research questionWhich results from simulation can be trusted on a physical robot, and how should simulation be built so that the difference is measurable?

Why it matters: Simulation makes it possible to test thousands of situations safely and cheaply. Its value depends on knowing where the model of the robot and its environment diverges from reality.
How we approach it: We run the same software interfaces in simulation and on hardware, vary physical parameters deliberately, and compare paired simulated and physical trials on the same tasks.
What progress looks like: A documented gap between simulated and physical performance for each task, and test scenarios that catch problems before hardware time is spent.

Manipulation and human–robot interaction

Scope: Object handling, task planning, operator interfaces and supervised interaction between people and robots.

Research questionHow can a robot handle varied objects and multi-step tasks while giving the people supervising it a clear view of what it is doing and an easy way to intervene?

Why it matters: Useful physical work usually involves picking, placing and adjusting objects, often near people. Operators need to understand and correct the robot without specialist knowledge.
How we approach it: We work on grasp and placement strategies, task sequencing and recovery, and on operator interfaces that show the robot’s state and intentions and record every intervention.
What progress looks like: Higher task completion across object variation, fewer and shorter interventions, and interaction logs that explain what happened in each trial.

Simulation render of a ROBOTIS OMY collaborative arm on a laboratory table with test objects. — Simulation render ROBOTIS OMY collaborative arm in an illustrative tabletop workcell. Simulation render by Axinome; not a photograph of an Axinome facility.

How the themes connect

The themes share one system. Perception provides the state that planning and learned policies act on. Simulation is where learned behaviour is trained and where perception and planning are tested against many scenarios. Manipulation and interaction are where the combined system meets objects and people.

Because the same software runs in simulation and on hardware, a finding in one theme can be checked in the others.

Illustration How hardware, software, simulation and evaluation relate in our work

Measuring the gap between simulation and hardware

Simulation is only useful if we know how far to trust it. We run paired trials, the same task with the same software in simulation and on the robot, and treat the difference as a quantity to measure and reduce.

Illustration Comparing simulated and physical trials

Application contexts

The research themes are chosen with practical uses in mind. The contexts below are potential applications: for each we describe the operational problem, the research challenge and the engineering approach we would take. None of them is presented as a completed customer project.

Potential application

Manufacturing and material handling

Operational problem: Parts and materials still move between stations by hand, especially where products vary and layouts change.
Research challenge: Handling varied objects reliably and recovering from small errors without stopping the line.
Proposed approach: Mobile manipulation with perception-driven grasping, task-level recovery and measured completion rates.

Potential application

Infrastructure inspection and condition monitoring

Operational problem: Plants, tunnels and utility sites require regular inspection rounds that are repetitive and sometimes hazardous.
Research challenge: Collecting consistent, comparable data on every round and detecting change rather than single readings.
Proposed approach: Repeatable mission routes on legged or wheeled platforms, calibrated inspection payloads and data pipelines that compare rounds.

Potential application

Warehouse navigation and logistics

Operational problem: Warehouses change daily: aisles are blocked, pallets move and people share the floor with machines.
Research challenge: Navigating safely through dynamic, partly unknown spaces and coordinating several robots.
Proposed approach: Localisation and planning that tolerate layout change, with fleet-level task allocation and traffic rules.

Potential application

Hazardous-environment observation

Operational problem: Some sites are unsafe or impractical for people to enter for routine observation.
Research challenge: Operating with limited communication and degraded sensing, while keeping a remote operator in control.
Proposed approach: Supervised autonomy with teleoperation fallback, robust telemetry and clear limits on autonomous behaviour.

Potential application

Research and education platforms

Operational problem: Laboratories and teaching groups need robots that students and researchers can program, test and compare.
Research challenge: Turning a capable robot into a well-documented, repeatable experimental setup.
Proposed approach: Software environments, simulation counterparts, example tasks and evaluation scripts built around a chosen platform.

Working on a related question?

We are interested in joint projects where a research question can be tested on real hardware.

Discuss a project