How robots see: cameras, lidar and navigation in crowds
Depth cameras, lidar and SLAM explained without jargon, including the documented weak spots (glass, featureless rooms, dense crowds) that matter when you plan a venue.
بقلم فريق تحرير humanoid4rent · نُشر في 7 نوفمبر 2026
Legged robots find their way using a mix of cameras, depth sensors and lidar. Software combines this data to build a map, work out where the robot is on that map and spot obstacles nearby. The approach works well in stable, well-textured rooms. Glass, plain walls, rooms that keep changing and dense crowds are still hard, so venue planning matters as much as the robot itself.
This explainer covers what the sensors do, how the software uses them, where the published research and manufacturer documentation say the system breaks down, and what that means for an event floor.
The sensors: what a robot actually carries
Most legged robots used at events carry at least two kinds of sensor. Each has its own strengths and its own blind spots.
Stereo cameras work roughly like human eyes. Two cameras a short distance apart see the same scene from slightly different angles, and software works out depth from the difference between the two images. Boston Dynamics states in its Spot developer documentation that the robot has five pairs of stereo cameras giving black-and-white images, with a 360-degree field of view, an operating range of 4 metres and a lighting requirement of above 2 lux. With a 4-metre range, the cameras cover the area close around the robot rather than a whole hall.
Lidar (light detection and ranging) sends out laser pulses and measures how long they take to bounce back. The result is a "point cloud": thousands of measured points that trace the shape of the room. Unitree's Go2 product page lists a "4D LiDAR L2" with a 360° × 96° hemispherical field of view and a minimum detection distance of 0.05 m, plus an HD wide-angle camera. A depth camera is an optional extra on the Edu version.
Humanoids often combine both kinds. On its G1 page, Unitree lists "Depth Camera + 3D LiDAR" as the G1's sensing package, alongside a four-microphone array. Unitree does not publish a field of view or detection range for those sensors there, so ask the supplier for those figures if they matter for your layout.
You can compare how individual models are equipped in the catalogue entry for the Unitree Go2 and other robot pages.
SLAM: building a map while finding yourself on it
Data from the sensors only becomes useful once software turns it into a sense of place. The standard technique is SLAM, short for simultaneous localisation and mapping. The robot builds a map of its surroundings and estimates its own position on that map at the same moment.
A widely cited survey in IEEE Transactions on Robotics, Past, Present, and Future of Simultaneous Localization And Mapping by Cadena and colleagues (2016), describes SLAM as exactly this joint problem. The authors also argue that robust performance over long periods, in changing environments, is still an open research question. In other words, SLAM works, but researchers do not regard it as finished.
Boston Dynamics' own navigation system for Spot, called GraphNav, shows how this looks in a commercial product. According to the Spot autonomy technical summary, an operator first drives the robot through the space by hand to record a map. The map is stored as a graph of waypoints, each holding a snapshot of what the sensors saw there. On later runs, the robot keeps track of its position relative to those waypoints.
From map to movement: obstacle avoidance
A map tells the robot where walls and fixed furniture are. A separate layer handles whatever is in the way right now: a stand that was moved overnight, a dropped bag, a person.
A common method is the costmap. This is a top-down grid in which each cell is marked as free, blocked or risky, and the path planner steers through the cheapest cells. NVIDIA's Isaac ROS nvblox documentation describes a GPU-accelerated package that rebuilds a 3D scene in real time from depth camera or lidar data and outputs a 2D costmap for the open-source Nav2 navigation stack. It also offers a mode that uses image segmentation to put people into a separate, people-only layer, so that moving humans are not mistaken for permanent walls.
That distinction matters at events. If a robot treats a crowd as fixed geometry, its map quickly fills up with phantom obstacles. If it ignores people, it may plan paths that are too close to them.
Where perception breaks down
The weak spots below are documented in research and manufacturer material. None of them is a defect of one brand. They follow from how the sensors work.
Glass and mirrors
Laser light mostly passes through clear glass or bounces off it at an angle, so the sensor may get no return at all, or a false one. The authors of a peer-reviewed paper in Sensors, LiDAR-Based Glass Detection for Improved Occupancy Grid Mapping (Tibebu et al., 2021), state plainly that most modern environments contain glass that is invisible to lidar. They propose a special filtering method to find it. Unless a robot runs that kind of extra processing, a glass balustrade, a shop window or a frameless display case may simply not appear on its map. Mirrors and polished chrome can also create reflections of rooms that do not exist.
Plain, featureless or changing spaces
Camera-based localisation needs visual detail to lock onto. Boston Dynamics warns in its GraphNav documentation that areas with sparse visual texture may not give the robot good features for tracking its location. It also warns that maps become unreliable when the environment changes significantly, and that the robot may become lost while navigating a recorded map. According to the same page, recovery from a lost state is not automatic: an operator has to step in and re-initialise.
For an exhibition hall this is significant. A map recorded during an empty build-up day may not match the same hall once stands, banners and visitors are in place.
Lighting
Spot's documented minimum is above 2 lux, which is very dim. That makes dark rooms less of a problem than many organisers assume. Strong contrast is a separate issue. Stage lighting, strobes, moving heads, haze and direct sun through a skylight can wash out or confuse camera images. Manufacturers rarely publish figures for these conditions, so ask the supplier how their particular robot has behaved under show lighting.
Dense crowds
Crowds are the hardest case, and the difficulty is as much about prediction as about sensing. A 2021 survey by researchers at the University of Washington, Carnegie Mellon University and Honda Research Institute, Core Challenges of Social Robot Navigation, describes the freezing robot problem. When a robot cannot predict how the people around it will move, it overestimates the uncertainty, decides that no safe path exists and stops. The authors add that the effect is not limited to big crowds: a single person close to the robot can already trigger overly evasive moves or a stop.
The same survey concludes that significant challenges still prevent autonomous robots from being deployed seamlessly in public pedestrian spaces. In practice, that is why robots at busy events are usually walked by an operator, kept to a defined route or placed in a managed zone, rather than left to roam freely. Our guide on teleoperated versus autonomous robots explains what audiences actually see in each mode.
Sensors and privacy
The cameras a robot needs to navigate are still cameras. Depending on the model and configuration, the robot may record, stream or process images of visitors, and lidar point clouds can capture the shapes of people in the space. Whether that counts as processing personal data depends on what is stored, for how long and who can access it. The supplier should be able to answer these questions precisely.
For events in the EU, our guide to cameras, microphones and visitor privacy and the article on what EU regulators say about video at public events set out the main questions. This article is general information, not legal advice.
What this means for event organisers
- Walk the route with the supplier. Point out glass walls, mirrored surfaces, balustrades and display cases before the event, and agree whether they need physical barriers or tape markings. The venue requirements checklist covers floor, space and access.
- Map the space as it will look on the day. If the robot uses a recorded map, ask whether it will be re-recorded after stands and dressing are in place, not just during an empty build-up.
- Expect stops in crowds and plan for them. A robot that halts when surrounded is behaving cautiously, not failing. Build in clear lanes, a defined performance area and an operator who can take over. The safety planning guide helps with zoning.
- Ask about show lighting. Share your lighting and haze plan with the supplier and ask how the robot copes with strobes, moving heads and direct sun.
- Clarify what the cameras keep. Ask whether images or point clouds are stored or streamed, where they go and when they are deleted, and reflect this in visitor information.
Frequently asked questions
What is the difference between lidar and a depth camera on a robot?
Lidar measures distance by timing laser pulses that bounce back from surfaces, producing a point cloud of the surroundings. A stereo depth camera uses two lenses and calculates distance from the small difference between the two images. Lidar is less dependent on visual texture, while cameras also capture what things look like. Many legged robots, such as the Unitree G1 according to Unitree's specifications, combine both.
Why do robots struggle with glass walls?
Laser light largely passes through clear glass or reflects away at an angle, so a lidar sensor may record no surface where the glass actually is. Peer-reviewed research describes glass as effectively invisible to lidar without additional processing. For events, this means glass balustrades, shop fronts and display cases should be pointed out to the supplier and, where needed, protected with physical barriers or markings.
Can a robot dog or humanoid walk through a crowd on its own?
Research on social robot navigation reports that dense crowds remain an open problem. Robots that cannot predict how people will move often slow down or stop, which researchers call the freezing robot problem. At busy events, robots are therefore usually guided by an operator, kept on a defined route or placed in a managed zone rather than left to move freely among visitors.
What is SLAM in robotics?
SLAM stands for simultaneous localisation and mapping. The robot builds a map of an unfamiliar space from its sensor data while estimating its own position on that map at the same time. It underpins most autonomous navigation, but survey research notes that keeping SLAM reliable over long periods and in changing environments, such as a hall filling up with stands and people, is still being worked on.
How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-11-07. Spotted an error? Write to info@humanoid4rent.com.
المصادر
جرى التحقق من الوقائع الواردة في هذا المقال بالرجوع إلى هذه الصفحات في يوم كتابته.
- Unitree Go2 – Unitree Robotics
- Unitree G1 – Unitree Robotics
- About Spot — Spot 5.2.0 documentation – Boston Dynamics
- Autonomy Technical Summary — Spot 5.2.0 documentation – Boston Dynamics
- Past, Present, and Future of Simultaneous Localization And Mapping: Towards the Robust-Perception Age – IEEE Transactions on Robotics (arXiv)
- Isaac ROS Nvblox – NVIDIA
- LiDAR-Based Glass Detection for Improved Occupancy Grid Mapping – Sensors (MDPI)
- Core Challenges of Social Robot Navigation: A Survey – arXiv
المزيد من المدونة
هل ترغبون في روبوت في فعاليتكم القادمة؟
صِفوا التاريخ والمكان وما يجب أن يقوم به الروبوت مرة واحدة فقط. يردّ الموردون بعروض أسعار يمكنكم مقارنتها على أساس متكافئ.
أرسلوا طلبًا