FIELD ATLAS
Six fields. A changing world.
Begin with your field of interest. Discover the problems robotics is working on and the stories behind them.
Editorial selection · source snapshot 23 Oct 2026
When you are unsure where to begin, start with a scene you recognize.
New possibilities in human spaces
Humanoids
Which tasks will bring robots into human workspaces first?
An impressive robot video can leave you hopeful and curious. Would it work at home, where objects rarely stay in exactly the same place?
A human-shaped body is only the beginning. Balance, manipulation and adaptation must work together in unfamiliar spaces. We look beyond an impressive demonstration to the tasks still unsolved and the trials taking place in real workplaces.
The challenge
Combining walking and manipulation while retaining performance on unfamiliar tasks. The range of repeatable work matters more than a single successful scene.
Where it leads
Follow factory pilots, object handling and research on locomotion and manipulation in domestic environments.
IN PRACTICE
What does the whole body do while carrying one part?
Boston Dynamics demonstrates Atlas moving engine covers from supply bins to a sequencing dolly. The task connects recognition, grasping and object-state estimation. These three stages explain how to read the demonstration.
Locate the grasp
Read the bin and part locations, then choose a hand pose. Atlas uses a vision model to locate fixtures and bins.
Move the body and object together
In general, lifting a load changes the forces on the body. Understanding a carry means considering the grasp together with the arms, trunk and legs.
Track the object until placement
Atlas combines vision, force and body-state sensing to follow changes. The demonstration also includes responses to unexpected events such as failed insertion.
WHAT THIS TELLS US
A carry is the complete journey from pickup to placement, not just the moment a hand lifts an object.
TWO WORDS TO KNOW
Two useful terms
- Whole-body control
- Coordinating hands, trunk and legs around the task and balance, rather than treating each body part independently.
- Proprioception
- Sensing the robot's own condition, such as joint position, joint speed and body tilt, alongside its view of the outside world.
What to look for
- Look for walking and handling combined into one task.
- Separate familiar scenes from previously unseen ones.
- LocoManiLab's 51.5 hours describe simulated teleoperation data.
The next generation of manufacturing
Industrial robotics
When does a good demonstration become dependable work?
Wanting less repetitive work and more time for meaningful tasks is familiar. Start with the specific work a factory robot is assigned.
Industrial robotics is about more than the speed of an arm. A machine must handle variation, adapt to changing processes and fit a workflow involving people. We follow the work that technology is being asked to do.
The challenge
Transferring learned skills to workplaces with different conditions and quality requirements. Transfer and verification after deployment need to be considered together.
Where it leads
Explore parts handling, assembly, internal logistics and laboratory automation, where repetitive work meets changing conditions.
IN PRACTICE
The robot that checks a connection after assembly.
In Universal Robots' Comprehensive Logistics case, a camera-equipped UR10 inspects connections in automotive subassemblies. Automation can check completed work before it moves on, as well as perform the assembly itself.
Approach the inspection point
The arm brings its camera to positions difficult for a fixed camera to see, including connections in confined spaces.
Image the connection
Images of the inspection points help assess whether connectors are seated and locked. Returning the camera to a useful view is central to the task.
Prepare for the next assembly
After imaging the inspection points, the arm retracts to its waiting position for the next assembly, connecting the inspection cycle to the line.
WHAT THIS TELLS US
Here, the value lies in repeatedly obtaining the view an inspection needs, rather than arm speed alone.
TWO WORDS TO KNOW
Two useful terms
- Machine vision
- Extracting task information such as position, shape or defects from images. Lighting and viewing angle also affect the result.
- Jigs and fixtures
- Devices that locate or hold parts, or guide a tool, helping establish repeatable conditions for the robot.
What to look for
- Name the task: AGCO plans parts sorting, stacking and tote placement.
- Distinguish a signed agreement, a pilot and reported results.
- Ask how the repeated task fits surrounding operations.
Autonomy meets everyday life
Mobility
When robots find a path, is the world ready to let them use it?
The journey matters as much as the destination. What does a robot need to notice when people and obstacles keep moving?
Knowing a destination and reaching it safely are different problems. Mobile robots must navigate people and moving obstacles while meeting the rules of the places they operate. We follow both the technology and the conditions for its use.
The challenge
Understanding planning in unfamiliar environments, the conditions for sharing space with people, and the rules supporting commercial use.
Where it leads
Our focus includes residential delivery, airports and rail facilities, construction sites and exploration in disaster environments.
IN PRACTICE
What if yesterday's open door is closed today?
Spot's Directed Exploration documentation describes a recorded patrol route becoming blocked. The robot can use its sensors to seek another path to a nearby waypoint instead of repeatedly approaching the same closed door.
Compare the route with the present
A route map records places and connections, but a door may have changed. Current obstacles must be considered alongside the recorded route.
Search for another connection
The documented system tries alternatives within the map before invoking Directed Exploration, which may find a nearby waypoint through another open door.
Rejoin the route
In the documented example, reaching a waypoint by the detour allows travel to the final destination. Avoidance is connected to resuming the original mission.
WHAT THIS TELLS US
Autonomous movement includes reconnecting a route through a changed environment while retaining the original goal.
TWO WORDS TO KNOW
Two useful terms
- Waypoint
- An intermediate reference location. A route can be represented by waypoints and the connections between them.
- Replanning
- Updating the route with new information when a passage is blocked or the estimated position changes.
What to look for
- Distinguish stationary obstacles from moving ones.
- Read the assumptions: SANDO bounds velocity, size and estimation error.
- Read simulations and physical flight experiments separately.
From sowing to weeding and harvest
Agriculture
How can robots take on farm work while protecting the crop?
A field looks different at sowing and as crops grow. Agricultural robots must work where needed while protecting the crop. Follow how planting records, soil conditions and growth stages shape their movements.
Agricultural robots work with the seasons and the condition of crops. They use recorded seed positions for weeding, locate fruit among leaves and travel along soil tracks and crop rows. Light, soil and growth stages change the conditions of otherwise repetitive work.
The challenge
Distinguishing crops from weeds, reaching occluded fruit and responding to changing ground and growth conditions. Read crop damage and human interventions alongside travel distance.
Where it leads
Explore mechanical weeding using seed positions, precision weeding with a quadruped, dual-arm apple harvesting and navigation within and beyond crop rows.
IN PRACTICE
Remember where seeds were placed, then weed between the plants.
FarmDroid's commercially available FD20 is a solar-powered seeding and weeding robot from Denmark. It records seed locations during sowing and uses them for subsequent mechanical weeding. This example follows the operating principles of a commercial machine, from planting to crop protection.
Record positions while sowing
FD20 uses RTK GPS to record seed locations. Because later work refers to this planting record, the relevant information is where seeds were placed, not only where the robot travelled.
Weed between rows and between plants
Tools work both between crop rows and between plants within a row. For the latter, a weeding knife moves in and out of the row. FarmDroid explains that known seed locations allow weeding before the seedlings become visible.
Leave room for the growing crop
A buffer around each seed coordinate keeps the weeding knife away. The manufacturer's guidance links this buffer and the tool's movement range to crop growth. The planting location stays fixed while later working conditions change.
WHAT THIS TELLS US
Agricultural automation involves more than completing one pass across a field. It connects the planting record with crop growth and later weed control.
TWO WORDS TO KNOW
Two useful terms
- Inter-row and intra-row weeding
- Inter-row weeding works between crop rows; intra-row weeding works between individual plants in the same row. The tool path and spaces to avoid differ accordingly.
- Crop buffer zone
- An area around a recorded seed position that the weeding knife must avoid. In FarmDroid's settings, its size is considered again as the crop grows.
What to look for
- Check which crops and planting methods the system is designed for.
- Look at how seed locations recorded during sowing guide later weeding.
- Consider how crop growth, soil and weather change the conditions for fieldwork.
Intelligence that sees, learns and acts
Robot AI
Can another robot use what one robot has learned?
Giving a good answer and doing a task well can feel like different abilities. Follow robot AI from a plan to its physical execution.
Robot AI goes beyond producing an answer on a screen. It must perceive objects, interpret contact and choose an action through a physical body. We follow how data and models become behavior.
The challenge
Checking whether a skill learned on one robot or in one setting transfers to another body and an unfamiliar situation. Generality must be read alongside its conditions.
Where it leads
Explore tactile manipulation, transfer across robot makers, handling deformable objects such as clothes and automating experiments.
IN PRACTICE
From a spoken request to a movement of the hand.
Google DeepMind's RT-2 studies a vision-language-action model trained with web data and robot demonstrations. The following explains how visual and language inputs connect to its action output.
Read the scene and instruction
Images show the available objects and an instruction specifies the task. Different instructions can select different targets in the same scene.
Represent the action
RT-2 represents actions such as position and rotation changes and gripper state as tokens, connecting a model's output to robot control.
Observe again after moving
Closed-loop control uses observations of the changed scene to continue acting, rather than replaying a complete motion from the initial image alone.
WHAT THIS TELLS US
Language understanding meets physical action here. Understanding a new instruction does not mean every new physical skill has also been learned.
TWO WORDS TO KNOW
Two useful terms
- VLA · vision-language-action model
- A family of models that takes images and language instructions and outputs robot actions, connecting interpretation to movement.
- Imitation learning and reinforcement learning
- Imitation learning learns behavior from demonstrations; reinforcement learning improves a policy using rewards following actions. A system can combine both.
What to look for
- Ask whether a failure starts in perception, planning or actuation.
- Alpha injects faults including sensor lag and grip slip in simulation.
- Check that conditions are comparable before reading success rates.
New missions beyond Earth
Space robotics
Will a movement learned on Earth work in space?
Familiar tasks can become surprisingly difficult in a new setting. Space robotics revisits grasping when the target is free to move.
Beyond Earth, familiar manipulation problems become new questions. Capturing an unsupported object requires reasoning about motion and contact together. We read research to distinguish current abilities from challenges that remain.
The challenge
Capturing freely moving objects in an unfamiliar environment. Results from a narrowly trained policy and a generalist must not be mistaken for a controlled, like-for-like comparison.
Where it leads
Begin with benchmarks for capturing objects inside spacecraft and manipulating them in space-related environments.
IN PRACTICE
Avoiding the next rock without waiting for Earth.
NASA describes Perseverance's AutoNav reading Martian terrain and planning around obstacles. The team on Earth plans the exploration route while the rover handles local driving decisions.
Send the exploration plan
The Earth-based team prepares a route using terrain images and science goals. Communication delay prevents driving as if the operator were beside the vehicle.
Interpret the local terrain
AutoNav builds a three-dimensional view of the terrain, identifies hazards and plans around them, reducing the need for a new instruction at each nearby rock.
Estimate the movement achieved
The rover also compares successive terrain images using visual odometry, helping distinguish wheel rotation from actual movement.
WHAT THIS TELLS US
Space autonomy can be understood as dividing distant planning from local decisions, rather than removing people from the mission.
TWO WORDS TO KNOW
Two useful terms
- Visual odometry
- Estimating movement from changes in terrain features across successive images.
- Communication delay
- The time a signal takes to travel a long distance, separating a command from the state report that returns.
What to look for
- Separate capture results for static and moving targets.
- Check how generalist and object-specific policies were trained.
- AstroBench's 2/200 and 8/10 come from separate experiments.