Signals InboxJuly 16, 2026Humanoid Robotics
Humanoids: what factory tasks can they do today?
Humanoids can already handle a narrow set of factory jobs today, especially rigid-part movement, fixture loading and simple inspection. The useful question is no longer whether they can work in factories, but which tasks survive contact with a real production line.
We track humanoid robotics daily. Want the market signals in your inbox?
Send me the signals →Humanoids can already perform real factory work today, but the proven jobs are still narrow: moving rigid components, loading parts into fixed fixtures, sequencing known objects and completing simple visual checks in structured automotive plants.
Only two of the 45 tasks we tracked reached the operational stage, while none qualified as scaled. Most factory activity still sits in pilots, repeated demonstrations or controlled workflows rather than broad industrial deployment.
The strongest tasks do not require human-like versatility. They work because the part, pickup point, route and destination barely change. Figure’s BMW deployment is the clearest example: impressive volume, but one tightly bounded job.
Assembly remains the weak point. Humanoids can place a badge or attempt a tolerant insertion, yet cables, seals, fastening, polishing and repair quickly expose the gaps in force control, tactile feedback and exception handling.
The real bottleneck is not walking or lifting anymore. It is completing long, error-sensitive sequences without a human stepping in when the part is misplaced, the tool slips or the production environment stops behaving exactly as expected.
Q1What factory tasks can humanoids do today?
Humanoid robots can already perform several factory tasks today, but the reliable work remains much narrower than most demonstration videos suggest.
The strongest evidence concerns moving rigid parts between known locations, loading components into fixed fixtures and completing simple inspection steps inside highly structured automotive plants. Once the work requires precise fastening, deformable materials, tool control or rapid recovery from unpredictable errors, the evidence becomes considerably weaker.
Humanoid factory tasks ranked by current capability
| # | Factory task | Category | Capability score | Evidence stage | What has actually been proven |
|---|---|---|---|---|---|
| 1 | Loading sheet-metal parts into fixed automotive assembly fixtures | Machine tending | 88 | Operational | Two Figure 02 robots performed this task during an 11-month deployment at BMW Plant Spartanburg in 2025. Figure reports more than 1,250 operating hours, over 90,000 parts loaded and a contribution to the production of more than 30,000 BMW X3 vehicles. The work was narrow and the fixtures were structured, but the duration and volume make it the strongest publicly quantified humanoid factory task we found. |
| 2 | Sequencing engine covers from supplier containers onto production dollies | Parts preparation | 72 | Piloted | Boston Dynamics has shown electric Atlas autonomously locating engine covers, lifting them from supplier containers and placing them into specified positions on a sequencing dolly. Atlas receives the required bin locations, detects the containers and parts through vision and corrects its movements without continuous teleoperation. The evidence is technically strong, although Boston Dynamics has not published production volumes comparable with Figure’s BMW deployment. |
| 3 | Inspecting installed door locks on automotive production lines | Quality inspection | 68 | Piloted | UBTECH’s Walker S participated in automotive quality-inspection training at Dongfeng Liuzhou Motor, including checking vehicle door locks. The robot operated on an actual assembly line, but no public cycle count, defect-detection rate or continuous runtime has been released. |
| 4 | Inspecting installed seat belts during final vehicle assembly | Quality inspection | 67 | Piloted | Walker S was also used to inspect seat belts at Dongfeng Liuzhou Motor. This provides factory evidence for a defined visual inspection step, although UBTECH has not disclosed how many vehicles were inspected, the level of human verification or the robot’s false-positive and false-negative rates. |
| 5 | Inspecting headlight covers for correct automotive installation | Quality inspection | 66 | Piloted | UBTECH reported Walker S inspecting headlight covers during training on Dongfeng’s assembly line. The task benefits from a stable inspection position and a limited number of expected configurations, but the published evidence does not establish production-scale reliability. |
| 6 | Applying vehicle badges to predefined exterior mounting positions | Product assembly | 65 | Piloted | Walker S has been shown affixing vehicle logos at Dongfeng Liuzhou Motor. This is a genuine assembly action rather than simple transportation, but it remains a relatively constrained placement task with a known object, surface and target location. No throughput or rejection-rate data has been published. |
| 7 | Carrying rigid automotive components between nearby factory workstations | Material handling | 64 | Piloted | Several industrial humanoid programs now target or test internal component transport. UBTECH says Walker S1 has entered vehicle assembly lines and can work with autonomous logistics vehicles, while Mercedes-Benz and Apptronik have tested Apollo for moving components to production workers. Public evidence confirms factory trials, but detailed volumes remain scarce. |
| 8 | Picking rigid parts from standardized factory storage containers | Parts preparation | 63 | Piloted | Atlas has autonomously localized and picked engine covers from supplier containers, while Figure demonstrated repeated handling of sheet-metal parts at BMW. The task is strongest when the container, part geometry and grasping area remain consistent. Performance on tangled, reflective or randomly piled components is less established. |
| 9 | Placing rigid parts into predetermined sequencing rack positions | Parts preparation | 62 | Piloted | Atlas has demonstrated autonomous placement of automotive components into assigned dolly positions. The robot can also recover from some failed insertions by searching for a dropped part, according to Boston Dynamics. Meaningful long-duration production data has not yet been published. |
| 10 | Walking between fixed stations while carrying a production component | Material handling | 61 | Piloted | Figure 02 walked an estimated 200 miles during its BMW deployment while moving parts around its workcell. Atlas and Walker robots have also demonstrated component transport in automotive settings. The task is feasible on prepared factory floors, although speed, human traffic and obstacle density can materially reduce reliability. |
| 11 | Loading and unloading standardized totes from line-side flow racks | Material handling | 60 | Piloted | Digit has repeatedly demonstrated tote handling, including lifting standardized containers and loading or unloading racks. Agility says Digit is commercially deployed across manufacturing, distribution and logistics, including with Schaeffler and Toyota Motor Manufacturing Canada. The clearest numerical evidence still comes from logistics rather than factory production, so the factory-specific score remains below the top tier. |
| 12 | Delivering assembly kits from staging areas to line-side workers | Material handling | 59 | Piloted | Apptronik describes Apollo’s near-term manufacturing role as material handling and line-side support. Agility and UBTECH also position their humanoids between logistics systems and production stations. The workflow is plausible and has entered factory programs, but public evidence rarely specifies completed kit deliveries or sustained cycle times. |
| 13 | Sorting known components into predefined production containers | Parts preparation | 58 | Piloted | UBTECH has deployed Walker robots for sorting work and coordinated multi-robot workflows in automotive factories. Sanctuary AI also reports testing its dexterous systems across hundreds of customer-defined tasks. However, task-level metrics separating autonomous sorting from supervised training remain limited. |
| 14 | Transferring parts between conveyors and adjacent production racks | Material handling | 57 | Piloted | This workflow combines abilities already shown separately by Figure, Atlas and Digit: identifying a standardized object, lifting it, walking a short distance and placing it in a known location. Factory pilots support its feasibility, but publicly documented continuous conveyor synchronization remains uncommon. |
| 15 | Performing repetitive pick-and-place beside a fixed assembly line | Parts preparation | 56 | Piloted | Figure’s BMW deployment provides strong proof for fixed-workcell pick-and-place, while UBTECH says Walker S can perform synchronized operations on assembly lines. Success is highest when object presentation and destination positions change very little between cycles. |
| 16 | Moving boxes or containers using a wheeled factory cart | Material handling | 54 | Repeated | Apptronik has consistently identified moving boxes and pushing carts as Apollo’s initial work. Digit has demonstrated similar bulk-material workflows. The motions have been repeated in controlled and customer environments, but detailed evidence from live manufacturing lines remains thinner than evidence from warehouses. |
| 17 | Collaborating with autonomous mobile robots during component delivery | Material handling | 53 | Repeated | UBTECH describes Walker S1 operating collaboratively with autonomous logistics vehicles inside vehicle-production environments. This proves basic system integration, although public information does not show how often the humanoid independently resolves delays, routing conflicts or incorrectly delivered materials. |
| 18 | Scanning component barcodes before completing a handling workflow | Parts preparation | 52 | Repeated | Boston Dynamics lists barcode scanning among Atlas’s workflow integrations. Combining scanning with transport is technically less difficult than dexterous assembly because the code and scanner position can be standardized. Customer-level accuracy and throughput data have not yet been published. |
| 19 | Replacing a depleted battery without ending the assigned factory workflow | Maintenance and support | 51 | Repeated | UBTECH markets Walker S2 around autonomous battery swapping, while the product version of Atlas is designed to exchange batteries rather than stop for charging. These systems address shift coverage, but public evidence of months-long autonomous battery management inside customer factories remains limited. |
| 20 | Feeding known components into a stationary processing machine | Machine tending | 50 | Repeated | Boston Dynamics identifies machine tending as a target Atlas application, and the underlying pick, carry and controlled placement actions have been demonstrated. However, published material provides less evidence for extended operation around active industrial machines than for parts sequencing. |
| 21 | Removing finished components from a stationary processing machine | Machine tending | 49 | Repeated | Removing a known part from a fixed machine is within the manipulation envelope already demonstrated by Atlas, Figure and several hand-focused systems. The weaker evidence concerns production integration, machine-state communication and safe recovery when parts are misaligned or unexpectedly hot. |
| 22 | Checking whether expected components are present at an assembly station | Quality inspection | 48 | Repeated | UBTECH’s factory inspection pilots and Apollo’s stated inspection workflows support simple presence-or-absence checks. These tasks are easier than detecting subtle defects because the robot is verifying a limited set of large visual features. Public accuracy data remains unavailable. |
| 23 | Separating correctly oriented parts from visibly misoriented components | Quality inspection | 47 | Repeated | Industrial humanoids can recognize known objects and poses, as shown by Atlas’s factory-part perception and UBTECH’s quality-inspection deployments. The task becomes less reliable when parts overlap, reflect light or differ only through small geometric details. |
| 24 | Recovering a dropped rigid component and restarting the placement attempt | Maintenance and support | 46 | Repeated | Boston Dynamics says Atlas can search for and retrieve a dropped factory part after an insertion failure. This is important because repeatability depends not only on successful cycles but also on recovering from ordinary errors. The behavior has been demonstrated, but its success rate across diverse objects is not public. |
| 25 | Packing rigid manufactured components into standardized shipping containers | End-of-line operations | 45 | Repeated | Sanctuary AI reports hundreds of tested customer-defined tasks, including manufacturing and logistics work, while Digit and Walker have repeatedly handled standardized containers. Packing becomes feasible when the products are rigid and their placement pattern is predefined. Mixed fragile products remain significantly harder. |
| 26 | Building a production order from several known rigid components | Parts preparation | 44 | Repeated | Boston Dynamics lists order building as an Atlas application. The robot’s sequencing demonstrations prove the central pick-and-place elements, but a complete multi-part order introduces more opportunities for identification, counting and placement errors. Customer production results are not yet public. |
| 27 | Inspecting large vehicle surfaces for obvious visible anomalies | Quality inspection | 43 | Repeated | UBTECH’s automotive programs demonstrate that humanoids can move around vehicles and inspect defined components. Large, visible anomalies are within reach, but detecting subtle scratches, paint defects or contour deviations at automotive quality standards requires better evidence and quantified detection performance. |
| 28 | Pressing large buttons or operating simple industrial control panels | Machine tending | 42 | Repeated | Humanoid hands and arms can physically press buttons and interact with controls designed for people. The challenge is not the contact itself but verifying machine state, selecting the correct control and responding safely when the equipment behaves unexpectedly. Public factory evidence is mostly demonstration-level. |
| 29 | Opening and closing lightweight machine doors with fixed handles | Machine tending | 40 | Demonstrated | Humanoid platforms have demonstrated doors and other contact-rich interactions, and machine tending is a declared Atlas application. Factory machine doors are more repeatable than ordinary doors, but reliable force control and confirmation that the door is fully secured remain insufficiently documented. |
| 30 | Placing small loose parts into predefined assembly-kit compartments | Parts preparation | 39 | Demonstrated | Dexterous humanoid hands can grasp and reposition small known objects, and Sanctuary has demonstrated advanced in-hand manipulation. The evidence does not yet show human-level speed or reliability when many visually similar small parts must be sorted without errors. |
| 31 | Inserting rigid components into moderately tolerant assembly openings | Product assembly | 38 | Demonstrated | Atlas has demonstrated controlled component placement and recovery after failed insertion. Rigid insertions with generous tolerances are achievable, but tighter fits require more tactile sensing, force control and correction than ordinary pick-and-place. |
| 32 | Rotating a component in one hand before final placement | Product assembly | 37 | Demonstrated | Sanctuary AI has demonstrated in-hand manipulation using hydraulic robotic hands. This is an important precursor to factory assembly because components are not always presented in a directly graspable orientation. Evidence is still primarily a capability demonstration rather than sustained production work. |
| 33 | Using tactile feedback to adjust grip on fragile manufactured parts | Product assembly | 35 | Demonstrated | Sanctuary has added tactile sensing and shown advanced hand control intended for delicate manipulation. Public demonstrations establish the underlying capability, but not reliable handling of fragile customer products over thousands of production cycles. |
| 34 | Tightening a clearly positioned fastener using a powered hand tool | Product assembly | 34 | Demonstrated | Humanoid systems can hold tools in controlled demonstrations, but robust industrial fastening requires alignment, reaction-force management, torque verification and safe recovery from cross-threading or tool slippage. We found no public evidence comparable in scale with Figure’s component-loading deployment. |
| 35 | Connecting rigid electrical plugs with generous alignment tolerances | Product assembly | 32 | Demonstrated | The action is possible under controlled conditions because it resembles a force-sensitive insertion. However, success depends on connector orientation, insertion force and verification of full engagement. Public humanoid factory evidence remains isolated and largely unquantified. |
| 36 | Handling flexible cables without tangling or damaging nearby components | Product assembly | 29 | Demonstrated | Flexible objects change shape during manipulation and cannot be localized like rigid parts. Current humanoids have shown increasingly dexterous hands, but we found no strong public evidence of sustained autonomous cable routing in a production factory. |
| 37 | Routing a wiring harness through several vehicle attachment points | Product assembly | 27 | Demonstrated | Wiring-harness installation combines deformable-object handling, obstacle avoidance, two-handed coordination and repeated force-sensitive insertion. Individual enabling abilities have been demonstrated, but the complete factory task remains far from operational evidence. |
| 38 | Applying adhesive along a precise three-dimensional production path | Component processing | 26 | Demonstrated | A Unitree G1 research deployment has been used to validate an automobile-window glue-application pipeline, showing that the task is technically possible with rapid object onboarding. This remains a research result rather than evidence of sustained autonomous factory production. |
| 39 | Performing consistent surface polishing with controlled contact pressure | Component processing | 23 | Demonstrated | Polishing requires continuous force control, stable tool orientation and compensation for changing surface geometry. Humanoid research has improved contact-rich whole-body control, but factory-quality consistency and long-duration tool use remain unproven. |
| 40 | Detecting subtle paint scratches under changing factory lighting | Quality inspection | 21 | Demonstrated | Humanoids can carry cameras and inspect predefined areas, but subtle cosmetic inspection requires controlled illumination, high-resolution sensing and validated defect-classification accuracy. Factory pilots have not publicly established competitive performance against specialized vision systems. |
| 41 | Installing flexible seals evenly around complex product openings | Product assembly | 18 | Claimed | This task combines deformable material, continuous force, two-handed coordination and verification that the seal is fully seated. Humanoid hand improvements make it a credible target, but we found no complete public factory execution with meaningful repeatability. |
| 42 | Starting and completing several different assembly tasks without reprogramming | Product assembly | 16 | Claimed | General-purpose humanoid companies frequently present task flexibility as their central advantage. Sanctuary reports testing 400 customer-defined tasks and Atlas is designed to learn new applications quickly, but switching autonomously between several real production jobs remains largely an ambition rather than an established operating mode. |
| 43 | Diagnosing an unexpected machine fault and selecting the correct intervention | Maintenance and support | 13 | Claimed | Current humanoids can perceive equipment, manipulate controls and follow learned procedures. Diagnosing an unfamiliar industrial fault requires causal reasoning, access to machine data and safe judgment under uncertainty. We found no evidence of autonomous production deployment for this complete task. |
| 44 | Repairing a production machine using several tools and replacement parts | Maintenance and support | 9 | Claimed | Tool handling, locomotion and visual perception have all advanced separately, but an autonomous repair combines diagnosis, tool selection, disassembly, part replacement, reassembly and validation. No current humanoid has publicly demonstrated this complete workflow at factory reliability. |
| 45 | Assembling a variable product from beginning to end at human speed | Product assembly | 5 | Claimed | No humanoid currently performs end-to-end variable-product assembly at human speed and reliability. Existing factory evidence is concentrated in individual handling, inspection and placement steps rather than full ownership of a changing production sequence. |
Q2Are humanoids actually working in factories today?
Humanoids are genuinely working in factories today, but only a small share of the tasks we tracked have moved beyond experiments and pilots. In our ranking, two of the 45 tasks reached the operational stage, 13 entered factory pilots, 13 were repeated in narrower environments, 11 were demonstrated and six remained principally claimed. None met our threshold for scaled deployment.
“Working in a factory” can describe very different realities. Figure 02 spent 11 months loading more than 90,000 sheet-metal parts at BMW’s Spartanburg plant. Other robots may enter a factory for a short test, collect training data or complete a filmed demonstration. Both get called deployments. They are not remotely the same thing.
The strongest quantified evidence still comes from a small number of programs. Figure provides the clearest production volumes. UBTECH covers a broader range of automotive tasks, including inspection, sorting and component placement, but usually without publishing comparable cycle counts or runtime. Boston Dynamics has shown technically strong autonomous parts sequencing, while Agility, Apptronik and others have entered manufacturing programs with less task-level operational data.
Humanoids have crossed the boundary from laboratory experimentation into real factory work. Broad industrial adoption is another boundary, and they have not crossed it yet.
Factory-task maturity across the 45 tasks tracked
| Evidence stage | Tasks in our ranking | Share of 45 tasks |
|---|---|---|
| Operational | 2 | 4% |
| Piloted | 13 | 29% |
| Repeated | 13 | 29% |
| Demonstrated | 11 | 24% |
| Claimed | 6 | 13% |
| Scaled | 0 | 0% |
Interested in humanoid robotics?We can send you all the signals
Send me the signals → Delivered straight to your inboxFigure AI launches Index with 16 million videos already uploaded
Chinese robot cut its 100m record from 9.39 to 8.86 seconds
XPeng’s robot unit raises $900M before selling a single robot
Unitree’s robot IPO jumps 460% in Shanghai
Unitree CEO sees robots hitting "a ChatGPT moment" by 2028
Unitree’s Superman robot hits 12.66 m/s, faster than Bolt’s peak
FCC blocks Unitree’s next robots, opening America’s humanoid race
San Francisco residents can now hire a humanoid cleaner for $30
Google DeepMind gives robots a Gemini brain that acts
Musk: it will be "America’s fastest industrial scale-up since WWII"
BotQ just built Figure’s 1,000th F.03 humanoid robot\
Tesla is building its first Optimus lines
Q3Are humanoids doing real assembly work yet?
Humanoids are doing a small amount of real assembly work today, but most activity described as manufacturing is still material handling around the assembly process. Only nine of the 45 tasks in our ranking involve directly modifying or assembling a product, and just one appears in the top 25.
The clearest example is UBTECH’s Walker S applying vehicle badges at Dongfeng Liuzhou Motor. The robot attaches a component to the product, so it qualifies as assembly. Still, the badge, mounting surface and target position can all be standardized. That is much easier than routing a wiring harness, connecting several plugs or tightening fasteners with verified torque.
Figure’s BMW deployment is operationally stronger but should not be misdescribed. Figure 02 loaded sheet-metal parts into fixtures and contributed to the production of more than 30,000 vehicles. It supported assembly; it did not independently assemble those vehicles. Boston Dynamics’ Atlas sequencing task sits even further upstream because the robot arranges parts for another worker or machine.
Humanoids are already creating value in factories, mainly as flexible interfaces between storage areas, fixtures, machines and human workers. General assembly workers, they are not.
We track humanoid robotics daily. Want the market signals in your inbox?
Send me the signals →Q4How autonomous are factory humanoids right now?
Factory humanoids can now execute narrow production cycles autonomously, but full-shift autonomy remains poorly proven. Companies frequently describe their robots as autonomous without disclosing how often humans intervene, reset the robot, approve a perception result or teleoperate an unusual recovery.
Boston Dynamics provides one of the clearest examples of narrow autonomy. Atlas receives bin-location instructions, detects containers and parts through vision, moves engine covers and can search for a dropped component after an insertion failure. Figure’s BMW deployment also strongly suggests substantial autonomous repetition because more than 90,000 parts were loaded over 1,250 operating hours.
A robot may execute the standard cycle alone while humans handle startup, replenishment, safety events, damaged parts and failed recoveries. Teleoperation may also teach the task or solve difficult exceptions without appearing in the final demonstration.
The industry has proved autonomy for repetitive, structured cycles far more convincingly than autonomy for complete shifts. The hard part is the abnormal case: the misplaced component, blocked route or failed grasp that cannot be allowed to stop production.
Q5Which factory jobs are humanoids still nowhere near doing?
Humanoids are still nowhere near autonomously owning complete assembly, maintenance or repair jobs. The bottom of our ranking is dominated by deformable materials, precise tool use, continuous force control, subtle quality judgments and workflows containing several dependent decisions.
Flexible cables, wiring harnesses and seals are particularly difficult because their shape changes during manipulation. Tightening a fastener requires alignment, torque control and confirmation that the connection is correct. Polishing and adhesive application require the robot to regulate pressure continuously while following a precise path. Maintenance adds another layer because the robot must diagnose the problem before selecting the correct intervention.
These are not necessarily the heaviest jobs. They are the jobs with the most uncertainty and the highest cost when something goes wrong. A dropped engine cover can usually be picked up again. A damaged connector, badly routed cable or poorly installed seal may create expensive rework several stations later.
The largest remaining gap is the ability to complete long, error-sensitive chains of actions while recognizing mistakes, recovering safely and maintaining human-level speed. Walking, lifting and basic grasping are no longer the main story.
We track humanoid robotics daily. Want the market signals in your inbox?
Send me the signals →We broke factory work into 45 distinct tasks across material handling, parts preparation, inspection, machine tending, assembly, processing and maintenance. The goal was to judge complete jobs, not infer readiness from a robot that can walk, grasp or perform one attractive motion in isolation.
For each task, we reviewed recent public deployments, pilots and demonstrations from Figure, Boston Dynamics, UBTECH, Agility Robotics, Apptronik, Sanctuary AI, Tesla and other humanoid developers. We separated documented production work from controlled demonstrations, research results and company claims.
The current capability score measures five elements: whether the complete physical action has been performed, how autonomous it was, whether it was repeated, whether it entered a real production workflow and whether it reached meaningful operational scale. Each element accounts for 20 points of the 100-point score.
Claimed means the task has been announced or targeted but not clearly shown from beginning to end. Demonstrated means it has been completed at least once under controlled conditions. Repeated means it has been performed multiple times with some evidence of reliability.
Piloted means the task entered a real factory or production-representative workflow. Operational means it was used repeatedly as part of actual production. Scaled requires substantial fleets, production volumes, shifts or replication across several sites.
Scaled is intentionally difficult to reach. Announced robot orders do not prove that those robots are successfully performing factory work, and autonomous execution of a normal cycle does not prove independent handling of setup, replenishment, safety events and unusual failures throughout a shift.
Where robots had demonstrated the individual abilities needed for a workflow but no developer had published the complete task, we treated that as evidence of technical feasibility rather than direct operational proof. That keeps inferred workflows below documented factory execution.
Key sources include Figure, BMW Group on humanoid robots in production, BMW’s Plant Spartanburg deployment note, Boston Dynamics on Atlas, Boston Dynamics’ electric Atlas update, UBTECH on Walker S, Agility Robotics on Digit, Apptronik on Apollo, Mercedes-Benz on its Apptronik pilot, Sanctuary AI’s technology material, Tesla’s AI and Optimus material, and research on humanoid adhesive application.
Building or investing in humanoid robotics?We can send you all the signals
Send me the signals → Delivered straight to your inbox