Source: https://broxme.com/research # Robotics intelligence & embodied AI. Our research focuses on how robots perceive their surroundings, make decisions, and act. We use indoor navigation to study perception, planning, and control. [See what we tested](https://broxme.com/research#research-evidence) [Read recent research](https://broxme.com/research#research-updates) Engineering evidence Controlled software tests Next milestone Controlled hardware trials planned Indoor navigation Blocked route ## Run the route Scripted simulation Start the virtual rover and watch it respond to a barrier. Ready at the start of the route. Enable JavaScript to run the route. Engineering evidence ## What we tested in simulation. These are observations from controlled tests in the Studio engineering build. Each test isolates one behavior with simulated inputs. The browser scene illustrates the same kind of problem using a scripted route. [Reproduce the browser example](https://broxme.com/robotics-development-guide#rover-example) . Route planning The A* planner found a route through an opening in a grid of obstacles and rejected a diagonal shortcut through a blocked corner. Obstacle response Simulated distance readings triggered a turn toward clearer space and a zero-motion command at the configured stop threshold. Localization Correcting the position estimate with simulated landmark readings reduced the final position error in a test with drifting wheel odometry. Line tracking The controller corrected its steering when it drifted off-center and entered search mode when the simulated line disappeared. Simulation and replay Identical inputs produced matching simulation and sensor outputs. Replay comparison detected a change when the simulated state differed. ## Research priorities. Indoor navigation gives us a shared test case for three questions about how robots behave in changing environments. ### Robot perception and planning Can a robot choose a useful next action when its route changes? We’re studying how to trace a decision from the sensor readings through the planner to the motion command. ### Embodied AI Can learning improve how a robot recovers when its original plan fails? We plan to compare a learned policy with our rule-based approach on the same scenarios. ### Indoor robot navigation Where does a response seen in simulation differ on a physical robot? A blocked aisle gives us a repeatable way to compare stopping, rerouting, and reaching a goal. Planned work ## From a virtual route to a physical test. We’ll start with one indoor mobile robot and one controlled course. The goal is to find which simulated behaviors carry over and turn the differences into repeatable tests. ### What we’ll measure - Goal completion across trials - Time and distance to stop - Time to resume movement - Operator interventions - ### Establish the simulation baseline Fix the start, goal, and obstacle location. Record what the robot senses, which route it selects, and what motion it commands in repeated runs. - ### Repeat the scenario on hardware Use the same course and compare the physical robot’s response with the simulated runs. Document where they agree and where the robot needs help. - ### Use failures to improve the next run Turn observed failures into regression tests. Use those scenarios to evaluate whether a learned recovery policy improves on the rule-based baseline. External research ## Recent developments in robotics AI. Recent work spans robot foundation models, simulation, and the benchmarks used to measure whether a learned skill transfers to a new task or robot. November 10, 2026 ### Gemini Robotics 2.5 Google DeepMind began rolling out its vision-language-action model to robotics partners. DeepMind reported evaluations across more than 200 tasks, including using tools and handling flexible or delicate objects. It also reported a pick-and-pack proof of concept at Mercedes-Benz using Apptronik’s Apollo 2. [Google DeepMind announcement](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/) November 7, 2026 ### Benchmarks for dexterity and adaptation At the Conference on Robot Learning, NVIDIA announced a developer preview of Isaac Lab-Arena 0.4 and highlighted two benchmarks, DexBench and VABAR. These benchmarks examine industrial manipulation and how robot policies adapt when the conditions or task change. [NVIDIA’s CoRL research updates](https://forums.developer.nvidia.com/t/news-nvidia-at-corl-2026-advancing-robot-learning/388260) October 22, 2026 ### Isaac Lab 3.0 NVIDIA released Isaac Lab 3.0 for general availability. The release expanded reinforcement-learning workflows and camera-based policy export. It also added configurable observation delay, which lets researchers study how a controller responds when sensor information arrives late. [Isaac Lab release announcement](https://github.com/isaac-sim/IsaacLab/discussions/8775) ## How to evaluate a learned robot policy. A vision-language-action (VLA) model combines camera input and a language instruction to predict robot actions. A foundation model can be trained across many tasks and adapted to a specific robot or workcell. Its performance in that setting depends on how well it handles new scenes, timing constraints, and recovery after a failed action. Generalization Keep layouts, obstacle placements, or instructions out of training and use them for evaluation. Compare the learned policy with a baseline using the same scenarios and starting conditions. Transfer to hardware Test delayed sensor readings, wheel slip, and changes in lighting in simulation first. On a controlled hardware course, measure completion, stopping distance, retries, and operator interventions for each scenario. Safety and recovery Define stop conditions before testing. Log when the robot stops, which event triggered the stop, and whether it can resume under supervision. Track safety interventions separately from task completion. Reproducibility Record the model and simulator versions, scenario settings, random seeds, hardware configuration, and timestamps. Keep traces of both successful and failed runs so another team can repeat the comparison. ## Technical references. External documentation and research for readers following indoor navigation, simulation, and embodied AI. - [Nav2 navigation in simulation](https://docs.nav2.org/jazzy/getting_started/quickstart/quickstart/) - [Gazebo simulation and sensor documentation](https://gazebosim.org/docs/latest/getstarted/) - [Google DeepMind’s Gemini Robotics paper](https://arxiv.org/abs/2503.20020) ## Bring us a navigation problem. If you build indoor mobile robots, tell us about a behavior you find hard to reproduce. We’re interested in working with teams on controlled tests. [Contact Broxme](https://broxme.com/contact?topic=research#write-to-us)