05 · The body
Sensors and perception
What a robot can measure, what each sensor lies about, and how software turns a pile of bad measurements into one usable picture of the world.
Every sensor on a robot is cheap, fast and wrong in its own way. A US$10 motion chip drifts by metres within a minute, a camera cannot tell a big far box from a small near one, and a laser scanner sees straight through a glass door. The robots that work in the field win by combining sensors whose failures do not overlap, and by keeping those sensors clean. For a company that services robots, the second half of that sentence is where the money is: most everyday "breakdowns" of a delivery robot are dirty or knocked sensors and a robot that has lost track of where it is.
Two kinds of sense
A robot senses two things: its own body and the world outside it. Your body does the same. Close your eyes and you still know where your hand is, because muscles and joints report their stretch. That inward sense is proprioception. Open your eyes and you see the cup on the table: that outward sense is exteroception.
Real
Proprioceptive sensors on a humanoid: an encoder on every motor and often a second one on the joint output, motor phase currents measured by the motor driver, joint torque sensors in some designs, an inertial unit in the torso, foot force sensors and battery telemetry. These run at 1 to 40 kHz and feed balance and joint control. Exteroceptive sensors: cameras, depth cameras, laser scanners, radar, ultrasonic sensors, microphones and fingertip touch. These run at 10 to 120 Hz and feed mapping, obstacle avoidance and the robot's AI model.
ELI5
Some sensors tell the robot "my knee is bent 40 degrees and I am tipping left." Others tell it "there is a chair two metres ahead." The first kind has to be very fast, because the robot falls over in a fraction of a second. The second kind has to be very rich, because the world is complicated.
The inward sensors live inside the joint module and are covered in Motors and joints. This chapter covers the inertial unit, foot force, and everything that looks outward. Touch sensing in the fingers is in Hands and touch.
Sensors cost less than you think and matter more than you think. In a Unitree G1 teardown the laser scanner and depth camera together cost about CNY 5,700 (about US$800) of a CNY 41,600 bill of materials, roughly 14% estimate, China Post Securities. Industry ranges put sensors at 10 to 25% of a humanoid's cost estimate. What they decide is whether the robot works outside the lab.
The inner ear: the IMU
An IMU (inertial measurement unit) is a chip that feels motion. It holds a gyroscope, which measures how fast the body is rotating, and an accelerometer, which measures push, including the constant 1 g push of gravity. At rest the accelerometer reads 1 g upward, which is how the robot knows which way is down. Every humanoid, quadruped and drone carries at least one, usually in the torso near the centre of mass.
Real
MEMS gyros (tiny vibrating silicon structures) measure angular rate through the Coriolis effect. To get position you integrate acceleration twice, so any constant error compounds: a 2 mg accelerometer bias (0.02 m/s²) becomes about 1 m of position error after 10 s and 36 m after 60 s, because error grows with time squared. Gyro bias is worse, since a small tilt error leaks gravity into the position estimate. The fix is architectural: use the IMU for fast, smooth motion and re-anchor it often with something absolute.
ELI5
Close your eyes in a moving car. You feel every turn and brake. After ten minutes you have no idea where you are, because every tiny misjudged nudge stayed in your running total. That growing error is drift. You fix it by opening your eyes now and then.
IMUs come in grades, and the price ladder is steep. Consumer chips such as the TDK InvenSense ICM-42688 and Bosch BMI088 cost US$1 to 10 and drift 1 to 10 degrees per hour. Industrial units like the Analog Devices ADIS16xxx family cost US$100 to 2,000 and drift 0.3 to 5 degrees per hour. Tactical and navigation grades, built for missiles and airliners, run from US$3,000 to US$250,000 order of magnitude. Robots use consumer and industrial MEMS. The spec sheet number is measured on a bench; on a walking robot, vibration and temperature swings dominate the real error, so good designs foam-mount the IMU and calibrate it warm.
Feet that feel the ground
A walking robot gets its best free fix from its own feet. When a foot is planted, it is not moving, so the software can reset the velocity estimate to zero on that leg. That trick, called a zero-velocity update, is why legged robots care so much about knowing the exact moment of foot contact. Designs sense contact three ways: estimating it from motor current and joint angles, adding a force sensor in each foot (Unitree's Go2 EDU has foot force sensors verified Sep 2026), or fitting a six-axis force/torque sensor at the ankle, which Chinese suppliers quote at CNY 5,000 to 10,000 per ankle estimate, 2025. Most cheap humanoids estimate; the sensor costs more than the rest of the leg's electronics.
Cameras: the richest sensor
A camera is the richest sensor per dollar by a wide margin. A US$10 to 100 module delivers millions of measurements per frame, and it is the only sensor that can read a label, recognise a face or see that a cup is half full. Its weakness is distance: a single image cannot tell a large far box from a small near one.
Four camera specs matter on a robot:
- Shutter. A rolling shutter exposes the image one row at a time, so fast motion smears and straight lines bend. A global shutter exposes every pixel at the same instant. Robots want global shutter, and RealSense's D430 series robotics cameras use it verified Sep 2026.
- Field of view. A fisheye lens sees 150 degrees or more at the cost of distortion that software must undo. 1X's NEO uses dual 8.85 MP stereo fisheye cameras at 90 Hz verified Sep 2026.
- Dynamic range. The span from dark to bright the sensor can hold at once. A robot rolling from a dim corridor into a sunlit lobby goes blind for a moment if this is poor.
- Latency. Time from light hitting the sensor to the software having the frame. Figure says Figure 03's new cameras doubled frame rate and cut latency to a quarter of Figure 02's verified. Every millisecond of perception delay is a millisecond the controller acts on old information.
Stereo: depth from two eyes
A depth camera reports a distance for every pixel. The most common kind is the stereo camera: two cameras a few centimetres apart, like your eyes. A point in the world lands in slightly different places in the two images. That shift, the disparity, gives the distance.
Real
Depth z = f × B / d, where f is the focal length in pixels, B the baseline between the cameras and d the disparity in pixels. Differentiate it and you get the rule that governs all triangulation: depth error grows with the square of distance. A RealSense-class unit with a 5 cm baseline is accurate to under a centimetre at 1 m, several centimetres at 3 m and close to useless at 10 m estimate. A wider baseline helps in proportion, which is why long-baseline models such as the D450 exist.
ELI5
Hold a finger near your nose and close each eye in turn: the finger jumps a lot. Hold it at arm's length: it jumps a little. Your brain measures that jump to judge distance. Far away the jump is so small that a tiny mistake in measuring it becomes a huge mistake in distance.
Stereo fails on blank walls and repeating patterns, where there is nothing to match or the wrong thing matches. Active stereo fixes blank walls by projecting its own infrared dot pattern, which washes out in sunlight. Two other depth methods matter:
- Structured light projects a known pattern and reads how it bends over surfaces. Face ID and the original Kinect work this way. It is precise and indoor-only.
- Time of flight (ToF) times light pulses: distance = speed of light × time ÷ 2. Indirect ToF, common in robot vacuums, is accurate to about a centimetre indoors out to about 5 m and is fooled by light bouncing off two surfaces. Orbbec's Femto cameras are ToF verified Sep 2026.
Two supplier facts shape the depth-camera market. Cognex announced on 22 Sep 2026 that it will buy RealSense, Intel's former depth-camera unit, for about US$500M, closing in Q4 2026 verified. The Unitree G1 base carries a RealSense D435i, costed at CNY 1,869 in a teardown estimate. And Orbbec (奥比中光) is headquartered in Nanshan, Shenzhen, with Gemini 330 stereo cameras and an IP67 Gemini 435Le verified Sep 2026. With RealSense moving into a US machine-vision group, Orbbec is the obvious non-US depth-camera partner for Chinese robot makers inference.
How many eyes named robots carry
| Robot | Cameras and depth | LiDAR | Other senses | Confidence |
|---|---|---|---|---|
| Unitree G1 | depth camera (RealSense D435i in teardown) | Livox MID-360 3D lidar | 4-mic array | verified |
| Unitree H2 | wide-field binocular camera | none listed | verified | |
| UBTech Walker S2 | RGB binocular stereo, learned depth | none in the head | verified | |
| 1X NEO | dual 8.85 MP 90 Hz stereo fisheye | none listed | 4 beamforming mics | verified |
| Figure 03 | head cameras (count not published; Figure 02 had 6 RGB) plus a palm camera in each hand | none listed | 3 g fingertip tactile sensors | count unpublished |
| Tesla Optimus | camera-only vision, like Tesla cars | none | joint force sensing | unverified |
| Boston Dynamics Atlas | 360° camera view | not published | tactile sensing | verified |
| Galbot G1 | head binocular, four torso RGB, wrist depth | chassis 3D lidar | 8 ultrasonic, IMU, two six-axis force sensors | unverified |
| Unitree Go2 | HD camera | L2 4D lidar | foot force (EDU) | verified |
| Unitree B2 | 2 depth, 2 optical | 1 3D lidar | IP67 body | verified |
| Pudu MP2000 | depth cameras | 3D lidar | 2,000 kg pallet mover | verified |
The pattern: Western humanoids and the newer Chinese full-size robots bet on cameras and learned depth, because their AI models learn from video and cameras are cheap. Robots that must navigate buildings reliably, from the G1 to every delivery and cleaning robot, still carry a laser scanner. Specs as of Sep 2026.
LiDAR: the laser tape measure
LiDAR fires laser pulses and times the echoes. Each echo is one exact 3D point. Sweep the beam and you get a point cloud, a pin-art sculpture of the surroundings. Where a camera guesses geometry, LiDAR measures it, accurate to about 1 to 3 cm per point estimate. It sees shape without meaning, though: a plastic bag and a rock are both "obstacle".
Scanners differ in how they steer the beam. Mechanical spinning units rotate a column of lasers on a motor for a full 360 degrees; bearings wear and shocks hurt. Semi-solid-state units steer one beam with a tiny mirror or rotating prisms (Livox uses prisms). Flash units light the whole scene at once and image it, with no moving parts and short range. Hemispherical units, built for robots, see a dome above and around the robot so a quadruped or humanoid can spot a step, a ledge and a low ceiling in one scan.
- Unitree L2
- 360° × 96°, 64,000 points/s, 30 m, 5 cm minimum range, 230 g, 10 W, built-in IMU, US$419 verified
- Hesai JT128
- 360° × 189°, up to 60 m, 1.152 million points/s, 265 g; price not listed verified
- Livox MID-360
- in the Unitree G1, CNY 3,840 in the teardown estimate; Livox is DJI's lidar arm in Shenzhen
- RoboSense E1R, Airy
- Shenzhen robot lidars; specs not verified in this research unverified
The price collapse
LiDAR is the clearest case of a sensor price falling a hundredfold. Velodyne's HDL-64E, the spinning bucket on every self-driving prototype from 2007 to 2017, cost about US$75,000 estimate. Car lidars at Chinese carmaker volume now run about US$200 to 500 estimate. Unitree sells its L2 robot lidar for US$419 verified Sep 2026. The spinning lidars inside Chinese robot vacuums cost about US$15 to 20 in parts estimate, and the vacuum industry made low-cost mapping ordinary before self-driving cars shipped.
| LiDAR | Era | Price | Confidence |
|---|---|---|---|
| Velodyne HDL-64E | 2007 to 2017 | ~US$75,000 | estimate |
| Robotics and survey pucks | 2025 | US$1,000 to 10,000 | estimate |
| Car lidar, Chinese volume (Hesai, RoboSense) | 2025 | US$200 to 500 | estimate |
| Unitree L2 | Sep 2026 | US$419 | verified |
| Robot vacuum lidar, parts cost | 2025 | US$15 to 20 | estimate |
A lidar that gets a hundred times cheaper can still bankrupt the companies that make it. About 70 venture-funded lidar start-ups existed around 2018. Ibeo went insolvent in Sep 2022, Quanergy filed for Chapter 11 by Dec 2022, and Velodyne merged into Ouster in Feb 2023, while Chinese volume makers Hesai and RoboSense won the car market control-systems primer. When a part deflates fast and its end market arrives late, the value flows past the part maker to whoever uses the part.
The humble sensors
Service robots rely on cheap sensors that never make a press release, and these are the ones that fail in daily use.
- Ultrasonic sensors send a sound pulse above human hearing and time the echo, like a car's parking sensor. They see glass and dark objects that lidar misses, at short range and with poor angular detail. Galbot's G1 is reported to carry eight unverified.
- mmWave radar uses radio waves, ignores fog, dust and darkness, and reads each target's speed directly from the Doppler shift. Car radars at 76 to 81 GHz cost about US$40 to 300 estimate. Its angular resolution is coarse, so it answers "something is approaching" better than "what is it".
- Cliff sensors are small downward-facing infrared or ToF sensors around the base that stop a wheeled robot before it drives off a step or down an escalator.
- Bumpers are the last line: a mechanical switch or pressure strip that stops the motors on contact. An emergency stop button cuts drive power outright; Pudu's manual lists a pressed or damaged e-stop as a standard reason a robot will not power on verified.
- Microphones for voice commands and speech. The G1 has a 4-mic array and NEO has 4 beamforming microphones verified. Pudu says its cleaning robots carry no microphones and keep camera frames only in runtime memory verified, which matters to buyers with privacy rules.
From pixels to a picture: the perception pipeline
Perception is the software that turns raw sensor data into facts the robot can act on: "there is a person 1.2 m ahead, moving left", "I am at table 12", "the cup's handle points at 40 degrees". It runs as a pipeline.
- Detection draws a box around each object and names it. Segmentation labels every pixel (floor, wall, person, cup). Pose estimation finds an object's position and orientation in 3D, or a person's skeleton, so a hand can grasp the handle or the robot can predict where a pedestrian is going.
- Mapping builds a map of the space, and localization finds where the robot is on that map. Doing both at once, in a new building, is SLAM.
- State estimation fuses everything into one best guess of the robot's own position, velocity and orientation, usually with a Kalman filter.
- Navigation and obstacle avoidance plan a path on the map and react to what moves.
Humanoids increasingly skip explicit steps: a vision-language-action model takes camera pixels in and produces joint targets out, with detection and pose learned inside the network. Wheeled service robots still run the classic pipeline, because it is predictable and certifiable. See Robot foundation models for the learned version.
SLAM: drawing the map while standing on it
Real
A SLAM front end turns raw sensing into relative motion: wheel and IMU odometry, lidar scan matching (aligning each new point cloud to the last), camera feature tracking, or visual-inertial odometry. It also proposes loop closures when a place looks familiar. The back end treats every pose as a node in a graph and every measurement as a spring, then solves for the most consistent arrangement (g2o, Ceres, GTSAM). Visual-inertial odometry drifts about 0.3 to 1% of distance travelled and lidar SLAM about 0.1 to 0.5% estimate, until a loop closure resets it.
ELI5
You wake up in a strange city with no map. You count steps and sketch as you walk, and your sketch slowly goes crooked. Then you turn a corner and see the bakery you passed an hour ago. Now you know your path is a loop, so you straighten the whole sketch in one go.
- Dead reckoning
The robot counts wheel turns and reads its IMU. Each small error stays in the total, so the estimated path (blue) ends away from the true start.
- Scans on a drifting path
Every lidar scan is placed where the robot thinks it was, so the map's walls smear and double.
- Loop closure
The robot recognises the start. That one match says the path must connect.
- Optimise
The solver bends every earlier pose a little so the loop closes, and the map snaps onto the real corridor.
SLAM is considered solved for structured indoor spaces with decent lighting and a world that does not change. It is still research for crowds, weather and places that look different every day contested. A restaurant is a changing world: chairs move, queues form, a wall of glass reflects the room. That is why delivery robots map once on installation day and then mostly localise against that saved map.
The Kalman filter and sensor fusion
Real
A Kalman filter alternates two steps. Predict: push the last estimate forward with a physics model and let its uncertainty grow. Update: blend in a new measurement, weighted by the Kalman gain, which trusts whichever source is currently less uncertain. Real robots use the extended Kalman filter (EKF), which re-linearises around the current estimate each step; it was developed to navigate Apollo and runs today inside every drone flight controller. Sensor fusion is the general practice: IMU for the fast signal, cameras, lidar or foot contact for absolute fixes, combined by an EKF or a factor graph.
ELI5
You track a friend through a crowd. Each second you guess where he has walked to, and your guess gets fuzzier. Then you glimpse him and correct your guess. If the glimpse was clear you trust it a lot; if it was blurry you trust your guess more. The Kalman filter is that habit written as maths.
Fusion only works if two plumbing jobs are done right. Calibration measures exactly where each sensor sits and how it distorts: the camera lens model, the camera-to-lidar offset, the IMU's bias at temperature. A camera mounted 2 mm off from its calibrated position puts every grasp 2 mm off. Time sync stamps every measurement with a shared clock, because a camera frame that arrives 30 ms late looks like the world moved. Engineers tune the filter's noise settings until the results look right, and every sensor knock or bracket swap in the field can void a calibration. That makes recalibration a real service line item.
The sensor suite of a service robot
A restaurant delivery robot carries a small, cheap, well-chosen sensor set. Pudu's BellaBot is the reference design: a 2D lidar window in the base for mapping and localisation, depth cameras higher up for 3D obstacles such as table edges and a child's head, touch sensing in its ears, a touch screen, and on older Pudu models a camera that reads markers on the ceiling. Wheel encoders and an IMU handle odometry.
- 1Cat-face touch screen
- 2Touch-sensing ears
- 3Four trays, 40 kg total
- 4Depth cameras for 3D obstacles 深度相机
- 5LiDAR window for mapping 激光雷达
- 6Controller, speakers, battery in the base
What goes wrong in the field
Sensors fail in the field in five repeatable ways, and each one is a maintenance task someone gets paid for.
| Problem | What happens | Maker's fix | Source |
|---|---|---|---|
| Dirty lenses | Fouled positioning sensor and 3D obstacle sensors cause "abnormal running"; grime makes a scrubber stop mid-route | Clean with lens tissue at least weekly, at once if soiled | Pudu manual verified; Gausium unverified |
| Glass and mirrors | Lidar beams pass through glass or bounce off mirrors and return from the wrong place | Tape or mark glass, add virtual walls to the map, fuse ultrasonic or camera | control-systems primer estimate |
| Reflective and dark floors | Shiny tiles act as mirrors to downward sensors; black surfaces absorb the laser and look like holes to cliff sensors | Adjust sensor thresholds, map no-go zones | control-systems primer estimate |
| Map drift and localisation loss | Older Pudu screens say "I'm lost, push me under the positioning marker"; moved furniture confuses the saved map | Push back to a marker; if self-test fails, call after-sales; remap after a refit | Pudu manual verified |
| Sunlight | Infrared dot projectors and indirect ToF wash out; cameras lose dynamic range at a sunlit entrance | Shade windows, route around, use lidar or passive stereo | control-systems primer estimate |
Two further causes show up in warranty terms. Unitree's warranty excludes damage from glass or icy floors and strong electromagnetic interference verified, and Pudu tells venues to protect the lidar and never lift the robot by its tray or head verified. A third-party restaurant-robot repairer says hardware faults cluster in cameras, sensors and calibration, and that hardware repair usually means swapping a module verified, Feb 2023.
Terms in this chapter
| English | 中文 + pinyin | ELI5 | Real definition |
|---|---|---|---|
| Sensor | 传感器chuángǎnqì | Any part that measures something and turns it into a number the computer can read: how far, how fast, how hot, how hard. | A transducer that converts a physical quantity (angle, acceleration, light, distance, force, sound) into an electrical signal sampled by the robot's electronics. Robots combine proprioceptive and exteroceptive sensors; sensors are about 10 to 25% of a humanoid's bill of materials (estimate). |
| Proprioception | The robot's sense of its own body: where each joint is, how hard each motor is pushing, which way it is tipping. Like knowing where your hand is with your eyes shut. | Sensing of the robot's internal state: joint angles from encoders, motor currents, joint torques, IMU orientation and rates, foot contact and battery telemetry. Sampled at 1 to 40 kHz and used by balance and joint control loops. | |
| Exteroception | The robot's sense of the world outside its body: seeing, hearing, feeling an obstacle. | Sensing of the environment: cameras, depth cameras, LiDAR, radar, ultrasonic sensors, microphones and tactile skin. Typically 10 to 120 Hz, feeding mapping, obstacle avoidance and learned policies. | |
| IMU (inertial measurement unit) | IMU / 惯性测量单元guànxìng cèliáng dānyuán | An inner ear on a chip. It feels turning and pushing, so the robot knows which way is down and how it is moving, but it slowly loses track of where it is. | A package of three-axis gyroscopes and accelerometers, usually MEMS, reporting angular rate and specific force at hundreds to thousands of hertz. Consumer MEMS parts cost US$1 to 10 and drift 1 to 10 degrees per hour; integrated position error grows with time squared, so IMUs are fused with absolute fixes. |
| Gyroscope | 陀螺仪tuóluóyí | The part of the IMU that feels spinning: how fast the robot is turning or tipping. | A sensor of angular rate. MEMS gyros measure the Coriolis force on a vibrating silicon structure; bias instability (degrees per hour of drift) is the headline spec. |
| Accelerometer | 加速度计jiāsùdùjì | The part of the IMU that feels pushes, including gravity pulling down, so the robot knows which way is up. | A sensor of specific force along each axis, including gravity (1 g at rest). Double integration gives position, so a 2 mg bias becomes about 36 m of error after 60 s. |
| LiDAR | 激光雷达jīguāng léidá | A laser tape measure spun around very fast. Each laser echo is one exact dot in 3D; millions of dots make a shape of the room. | Light detection and ranging: pulsed-laser time of flight swept by a spinning head, mirrors, prisms or a flash array, giving 10^5 to 10^7 points per second at about 1 to 3 cm accuracy. Blind to glass, mirrors and very dark surfaces. Robot units run from about US$419 (Unitree L2, Sep 2026) upward. |
| Depth camera | 深度相机shēndù xiàngjī | A camera that tells you how far away every pixel is, not only its colour. | A sensor producing a per-pixel range image by stereo matching, structured light or time of flight. Examples: RealSense D435i (in the Unitree G1 base), Orbbec Gemini 330 and Femto series. |
| Stereo camera | 双目相机shuāngmù xiàngjī | Two cameras side by side, like your two eyes. The closer something is, the more it shifts between the two pictures, and that shift gives the distance. | A pair of calibrated cameras a baseline B apart; depth z = f × B / d from disparity d. Depth error grows with the square of distance. Walker S2 and Unitree H2 use binocular stereo instead of lidar. |
| Structured light | 结构光jiégòuguāng | Shine a known pattern of dots or stripes onto things and watch how it bends over their shape. | Active depth sensing that projects a known infrared pattern and triangulates its distortion. Precise at short range indoors; washes out in sunlight. Used in Face ID and the original Kinect. |
| Time of flight (ToF) | ToF / 飞行时间fēixíng shíjiān | Flash light and time how long it takes to bounce back. Light is so fast the stopwatch has to count in billionths of a second. | Range measurement from the round-trip time of light, distance = c × t / 2. Indirect ToF measures phase shift of modulated light (about 1 cm indoors to about 5 m, fooled by multipath and sunlight); direct ToF times photons with SPAD detectors and underlies lidar. |
| Point cloud | 点云diǎnyún | Thousands of dots floating in 3D, each one a spot a laser or depth camera touched. Together they show the shape of things. | A set of 3D points, often with intensity and time stamps, produced by lidar or depth cameras. It carries geometry without meaning; segmentation and detection add labels. |
| mmWave radar | 毫米波雷达háomǐbō léidá | Radar with very short radio waves. It sees through fog and dust and can tell how fast something is coming, but the picture is blurry. | Millimetre-wave FMCW radar (76 to 81 GHz in cars) that measures range and, through Doppler, radial velocity per target. Weather-proof and cheap (about US$40 to 300); angular resolution is coarse unless imaging arrays are used. |
| Ultrasonic sensor | 超声波传感器chāoshēngbō chuángǎnqì | A tiny speaker and microphone that squeak too high for you to hear and listen for the echo, like a car's parking sensor. | A short-range acoustic time-of-flight sensor. It detects glass and dark objects that lidar misses, with a wide beam and poor angular resolution; common around the base of service robots. |
| Cliff sensor | 防跌落传感器fáng diēluò chuángǎnqì | Little eyes under the robot that look at the floor and stop it before it rolls off a step. | A downward-facing infrared or ToF range sensor on a wheeled base that detects drops such as stairs and escalator edges. Very dark or shiny floors can trigger false cliffs. |
| SLAM | SLAM / 即时定位与地图构建jíshí dìngwèi yǔ dìtú gòujiàn | Drawing a map of a new building while also working out where you are on it, and fixing the map when you recognise a spot you have seen before. | Simultaneous localization and mapping: a front end (odometry, scan matching, feature tracking, loop-closure proposals) plus a back end that optimises a pose graph (g2o, Ceres, GTSAM). Lidar SLAM drifts about 0.1 to 0.5% of distance and visual-inertial odometry about 0.3 to 1% between loop closures (estimates). |
| Localization | 定位dìngwèi | Working out where you are on a map you already have. | Estimating the robot's pose against a stored map by matching current sensor data to it. Service robots map once on installation and localise every run; 'localization lost' (定位丢失 dìngwèi diūshī) is a standard field fault. |
| Mapping | 建图jiàntú | Drawing the map of a place so the robot can find its way next time. | Building a persistent map (occupancy grid, point cloud or feature map) of an environment, usually on installation day for service robots, and redoing it after a refit. |
| Navigation | 导航dǎoháng | Planning a route from here to there on the map and following it without bumping into anything. | Global path planning on a map plus local planning around obstacles, combined with localization. In ROS 2 this is typically the Nav2 stack; service robots add lift and door integration. |
| Calibration | 标定biāodìng | Measuring exactly where each sensor sits and how it is slightly off, so the robot can correct for it. A camera moved by 2 mm makes every grasp miss by 2 mm. | Estimating intrinsic parameters (lens model, IMU bias, scale factors) and extrinsic parameters (the pose of each sensor relative to the body and to each other). Knocks, bracket swaps and temperature can void it, so field recalibration is a service task. |
| Sensor fusion | 传感器融合chuángǎnqì rónghé | Combining several sensors that each make different mistakes, so together they are right more often than any one alone. | Combining measurements from sensors with complementary failure modes (IMU with camera, lidar and foot contact) into one estimate, usually with a Kalman-family filter or a factor graph. Requires calibration and time synchronisation. |
| Kalman filter | 卡尔曼滤波kǎ'ěrmàn lǜbō | A habit written as maths: guess where something is now, then correct the guess with each new measurement, trusting whichever is less fuzzy. | A recursive estimator that alternates prediction with a process model and update with measurements weighted by the Kalman gain. The extended Kalman filter (EKF) linearises nonlinear models and runs in drone flight controllers, robot state estimators and battery management systems. |
| Obstacle avoidance | 避障bìzhàng | Seeing something in the way and going around it or stopping. | Local reactive planning that slows, stops or re-routes the robot around detected obstacles from lidar, depth cameras, ultrasonic and bumper inputs. A core selling point and a common complaint in service robots. |
| Perception | 感知gǎnzhī | The software that turns camera pictures and laser dots into facts: that is a person, that is a cup, I am at table 12. | The pipeline from raw sensor data to a world model: detection, segmentation, pose estimation, mapping, localization and state estimation. In humanoids it is increasingly learned inside a VLA; service robots still run classic modules. |
| Global shutter | A camera that takes the whole picture at the same instant, so fast motion does not smear or bend. | An image sensor that exposes all pixels simultaneously, avoiding the rolling-shutter distortion of row-by-row exposure. Preferred for robots and used in RealSense's D430 series. | |
| Loop closure | 回环检测huíhuán jiǎncè | The moment the robot recognises a place it has been before and uses that to straighten its whole map. | In SLAM, a detected revisit of an earlier place that adds a constraint between distant poses; graph optimisation then removes accumulated drift. False closures from look-alike places bend the map. |
Sources for this chapter
- Unitree product pages: G1 unitree.com/g1, H2 unitree.com/H2, Go2 unitree.com/go2, B2 unitree.com/b2, L2 lidar unitree.com/L2
- Hesai JT128: hesaitech.com
- RealSense, Cognex to acquire RealSense, 22 Sep 2026: realsenseai.com
- Orbbec: orbbec.com
- China Post Securities G1 teardown via chinabizinsider, 29 Apr 2026: chinabizinsider.com
- Figure, Introducing Figure 03: figure.ai
- 1X NEO specifications: 1x.tech/neo
- UBTech Walker S2: ubtrobot.com
- Boston Dynamics Atlas: bostondynamics.com
- Pudu robot manual (FCC filing): fcc.report; Pudu support FAQ: pudurobotics.com
- Gausium scrubber maintenance, 24 Jun 2026: gausium.com
- Back of House, restaurant robot repair: backofhouse.io
- Groundcrew research dossiers, Sep 2026, built from the sources listed above.