{"schema_version":"1.0","service":"Publicasta","type":"article","id":724,"slug":"iros_2026_digit_mobile_manipulation_evidence","title":"IROS 2026’s Digit Demo Shows the Next Humanoid Test Is Measurable Reliability","excerpt":"A first-try mobile-manipulation run on Agility’s Digit v4 is a meaningful research signal, but not proof of general warehouse autonomy. The real test is repeatability, recovery, safety and cost in a working facility.","language":"en","default_language":"en","canonical_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=en","image":{"url":"https://publicasta.com/storage/projects/11/pages/724/2026/09/fb31021c-090b-43ee-9b17-9b5f342bab55.webp","alt":"A bipedal humanoid warehouse robot uses its arms to grasp a tote while moving through an industrial facility."},"publisher":{"id":11,"slug":"robots","name":"Robots and Autonomous Systems","url":"https://publicasta.com/robots"},"author":{"name":"Anton R"},"published_at":"2026-09-28T17:25:35+00:00","updated_at":"2026-09-28T17:25:35+00:00","content_markdown":"The most useful humanoid-robot claim from IROS 2026 is also the easiest to overread. A short conference-floor report says an Agility Digit v4 completed an end-to-end mobile-manipulation run on its first attempt. That is interesting because it joins the two jobs that usually get demonstrated separately: moving through a space and doing useful work with the arms. It is not, by itself, evidence that a humanoid can reliably run an entire warehouse shift.\n\n ![A bipedal humanoid warehouse robot uses its arms to grasp a tote while moving through an industrial facility.](https://publicasta.com/storage/projects/11/pages/724/2026/09/fb31021c-090b-43ee-9b17-9b5f342bab55.webp)\n\n The distinction matters. Robotics companies are moving from carefully staged demonstrations toward systems that must navigate, perceive, grasp, recover from mistakes and remain safe around equipment and people. A single successful rollout can show that a controller, robot and task were compatible on that occasion. It cannot establish a production rate, a safety case, a failure distribution or a cost advantage.\n\n IROS 2026, held in Pittsburgh from September 27 to October 1, is a good place to examine that gap. Its program puts mobile manipulation, whole-body control, learning and physical safety beside one another. The conference-floor Digit report supplies the news hook. Agility’s own disclosures supply the commercial context. Together they point to a less dramatic but more useful conclusion: the next test for humanoids is not whether a learned policy can make a robot move its body and hands at once. It is whether that capability can be measured, bounded and operated as a service.\n\n ## What the IROS demonstration actually says\n\n The public claim came from Chris Paxton, who described an end-to-end mobile-manipulation run on an Agility Digit v4 at IROS 2026 as working “out of the box” and on the first try. A robotics-news account that preserved the post and the available clips adds an important qualification: the report did not include a success rate, a number of attempts, a task list or a controlled comparison.\n\n That is still a meaningful result. In a conventional robotics stack, navigation, arm motion, grasping and task logic may be separate modules. They can pass state between one another, but each handoff is a place where assumptions become brittle. A mobile manipulator may arrive at an object with a slightly wrong orientation; an arm planner may expect the base to be stationary; a grasp may fail because the object is not where the camera model predicted. “End-to-end” generally signals that a learned or jointly controlled system is coordinating more of the robot’s behavior as one decision problem.\n\n The phrase does not settle how much of the system is learned, how much is engineered, or how the task was selected. It also does not tell us whether a person was available to intervene, whether the route and object arrangement were prepared, or whether the robot had seen the scene during prior data collection. Those are not accusations. They are the basic variables needed to interpret a demonstration.\n\n A first-try success can therefore support a narrow statement: Digit v4 was able to execute at least one integrated mobile-manipulation scenario under the conditions of the demonstration. It does not support the broader statement that Digit can generalise to arbitrary warehouse work. The useful editorial question is not whether the clip is real. It is what evidence would turn the clip into an operational claim.\n\n ## Why mobile manipulation is harder than a pick-and-place clip\n\n A stationary arm has a major advantage: its workspace, fixtures, camera position and collision boundaries can be engineered in advance. Industrial cells often look repetitive because repeatability is part of the product. The robot is not asked to understand an entire building; it is asked to perform a defined motion inside a known envelope.\n\n A mobile manipulator gives up some of that control. Its base changes the geometry of every reach. The floor can be uneven. People can cross the path. A tote can be partly obstructed. The robot must decide where to stand, how close it can safely approach, whether it can reach from that position and what to do if the first grasp fails. If it is bipedal, maintaining balance while reaching or carrying introduces another coupling between locomotion and manipulation.\n\n That coupling is why the word “mobile” carries so much weight. Walking to a marked point is navigation. Picking an object from a fixed presentation is manipulation. Doing both while adapting to the actual position of the object is a longer-horizon capability. The action space expands, and so does the number of ways a small error can propagate.\n\n The IROS program reflects this research problem. The conference lists sessions on reinforcement learning for agile humanoid and legged locomotion, difficult-ground locomotion and work on communication and collaboration. A dedicated Mobile Manipulation and Embodied Intelligence workshop describes contact-rich mobile manipulation as a foundation for robust, deployable autonomy. That framing is more informative than a general promise of “general-purpose intelligence”: contact, recovery and deployment conditions are where a robot’s assumptions meet the physical world.\n\n ## Digit’s commercial baseline is already narrower than the headline\n\n Agility’s public materials place Digit v4 in a specific industrial role. In a filing made in connection with its proposed business combination, the company describes v4 as deployed, with a stated carrying capacity of up to 35 pounds, a claimed four-hour runtime and autonomous charging. The same presentation says the robot has 360-degree awareness, adaptive dexterity aimed at industrial workflows and whole-body control that includes vision-language-action models.\n\n Those are company claims and should be read as such. The filing also shows the intended work: a Digit v4 placing 25-pound baskets of bearing components from a stamp press into an industrial washing machine. This is not a claim that the robot can perform every task a human can perform. It is a concrete material-handling workflow with a defined object class, route and destination.\n\n Agility’s September 2026 press release introduces Digit v5 as a later generation designed for cooperative work at scale. The company’s presentation contrasts v5’s planned capabilities with v4, including a larger stated lifting capacity, longer operating potential, greater reach and changeable end effectors. It also describes v5 as intended for operation outside a workcell and alongside people. “Planned” is the important word: those v5 figures are expectations until the product is released, tested and deployed under the stated conditions.\n\n This commercial baseline makes the IROS run easier to interpret. The demonstration may represent an expansion from a purpose-built material-handling workflow toward more integrated control. But it should not be confused with the company’s already disclosed deployment evidence, and neither should be confused with a general warehouse substitute. There are at least three different milestones:\n\n - a robot can complete a selected task in a demonstration;\n- a robot can repeat a defined workflow with a measured intervention rate;\n- a robot can deliver acceptable economics and safety over a real operating schedule.\n\n The first is a research result. The second is an engineering result. The third is a business and operations result. Humanoid coverage often compresses them into one sentence.\n\n ## The denominator is the missing part of most robot demos\n\n “First try” sounds precise but has no denominator. First try out of how many prepared trials? Was the run selected because it succeeded? How many failed attempts occurred during setup? What counted as success: touching the object, grasping it, placing it in the correct location or finishing the complete sequence without human assistance?\n\n A serious evaluation would publish the task definition and the conditions around it. At minimum, readers need to know:\n\n - how many runs were attempted and how many succeeded;\n- whether the object, route and destination changed between runs;\n- whether the robot had prior demonstrations or scene-specific tuning;\n- whether a teleoperator or safety observer intervened;\n- the duration of each run and the time spent recovering;\n- the consequences of a failed grasp, collision or navigation error;\n- what sensors and compute were active on the robot;\n- whether the same policy ran on a second robot or a second site.\n\n This is not bureaucracy for its own sake. It is how a capability becomes comparable. A 95 percent success rate on a fixed, clean presentation may be excellent for one workflow and useless for another. A 70 percent rate in a cluttered environment might be a valuable research step if the robot can detect uncertainty and ask for help safely. Without the denominator and operating conditions, a video mostly measures the quality of the selected moment.\n\n The same logic applies to autonomy. A robot can be autonomous in low-level motion while still relying on a person to choose the task, approve a recovery or reset an object. Human oversight does not make a system fraudulent; it changes the service being sold. A warehouse operator buying a robot that works with occasional assistance is buying a different product from one that runs unattended.\n\n ## Why integrated control is still a real technical advance\n\n The caution should not hide the research value. Coordinating locomotion and manipulation is difficult because the robot has to maintain a useful body configuration while its contact points and balance are changing. A base motion that improves reach can reduce stability. A cautious stance can make the arm miss the object. A grasp can change the load distribution. If the controller treats each stage independently, it may produce locally sensible actions that fail as a sequence.\n\n An end-to-end policy can, in principle, learn relationships across those stages. It may learn that a better approach angle reduces grasp failures, or that the base should stop earlier because the arm needs room to close around the object. With enough varied data and a suitable safety layer, this can reduce the brittle interfaces between hand-written planners.\n\n But end-to-end does not mean magic. A policy still depends on the observation space, training distribution, action limits, hardware condition and recovery design. A learned system can be smooth and fast in familiar situations while being poorly calibrated at the edge of its experience. The controller also needs a way to recognise when it does not know what to do. For physical machines, uncertainty is not merely a confidence score on a screen; it can become a dropped load, an unexpected movement or a person being forced to step into the work area.\n\n That is why the best near-term architecture may be hybrid. Learned components can propose movements and interpret scenes. Classical control, collision monitoring, force limits, geofencing and supervisory logic can constrain what happens next. The exact split will differ by task, but commercial deployment requires more than a policy that looks capable in a clip.\n\n ## The operational test is recovery, not just completion\n\n A warehouse does not pay for a robot to succeed once. It pays for completed work over time. The expensive events are often not spectacular failures; they are the small interruptions that require an employee to stop, clear an object, re-stage a tote, reboot a system or explain why the robot has become cautious.\n\n A deployment report should therefore include recovery metrics alongside task success. How often does the robot pause? How long does it take to resume? Can it safely return to a known state after a failed grasp? Does it identify the cause of the failure, or does a person have to inspect the scene? What happens when a battery, gripper, camera or network connection degrades?\n\n Agility’s filing says customer deployments had accumulated more than 65,000 hours of operation and describes deployments or commitments across nine customer facilities. Those numbers are more relevant to maturity than a single stage run because they point toward repeated use in real environments. Yet even operating hours need context. The public material does not, by itself, provide a complete breakdown of productive hours, interventions, uptime, task mix, labour savings or cost per completed move.\n\n The right reading is neither dismissive nor credulous. The hours indicate that the company has moved beyond a lab-only story. They do not prove that every workflow is economical, that every site has the same performance or that the robot can operate without support. Real deployment is evidence of engineering and customer willingness. It is not automatically evidence of positive unit economics.\n\n ## Safety changes when the robot leaves the workcell\n\n A robot inside a fenced or otherwise controlled cell can be designed around predictable boundaries. A robot operating beside workers needs a different safety case. It must limit force and speed, detect people and obstacles, signal intent, stop safely and restart in a controlled way. Its software updates, maintenance procedures and exception handling become part of the safety system.\n\n Agility’s materials describe Digit v4 as having safety systems and a workcell-oriented deployment path, while the company positions Digit v5 as intended for cooperative operation around people. That progression is plausible as a product strategy, but it also shows why the generations should not be treated as interchangeable. A future design target is not a present certification. A marketing diagram is not a hazard analysis.\n\n The physical environment matters too. A robot that is safe around a marked pallet lane may not be safe near a worker carrying a long object, a forklift turning blind or a wet floor. A successful manipulation run says little about these interactions unless they were part of the test. The more general the environment, the more the safety argument has to cover unusual but foreseeable situations.\n\n For buyers, the practical question is not simply “Is the robot safe?” It is “Which hazards are controlled by the robot, which by the facility and which by human procedures?” That question leads to specific requirements: restricted zones, speed limits, emergency stops, inspection schedules, operator training, incident logging and a clear path for disabling the machine. Autonomy is a system property, not a label attached to the model.\n\n ## Cost and maturity: humanoid shape is not the same as value\n\n A humanoid form can be useful where the workplace is already built for people. Two legs can use existing aisles and stairs; an upright body can reach shelves and interfaces designed for human hands. That flexibility is the argument for general-purpose form factors.\n\n It is also a cost. A biped needs balance control, fall management, more complex maintenance and a larger safety envelope than a fixed arm or specialised wheeled machine. If the task is moving a tote along a flat, predictable route, a conveyor, cart or mobile base may be cheaper and easier to keep running. A humanoid earns its complexity when its ability to use existing spaces and tools offsets the cost of that complexity.\n\n The business model matters as much as the purchase price. Robot-as-a-Service can move the decision from capital expenditure to an operating contract, but it does not make labour disappear. The customer still pays for floor space, integration, charging, supervision, maintenance and the cost of exceptions. A robot that is inexpensive per month but requires frequent intervention may not beat a specialised machine. A more expensive system could make sense if it covers several workflows without a major reconfiguration.\n\n This is where Digit’s current profile is more credible than broad “general-purpose robot” language. A defined industrial workflow gives the supplier a chance to measure output and build the support process. The product can expand later if the data shows that the next task is worth solving. The commercial path is likely to look less like a human replacement arriving all at once and more like a sequence of bounded jobs connected by a growing control and operations platform.\n\n ## What to watch after the conference\n\n The next useful disclosures will be less cinematic than the IROS clip. Look for repeated trials, task descriptions, intervention rates and evidence from a second site. Look for the difference between a policy that completes a route and one that can recover from an object being moved, a person entering the path or a gripper losing its hold.\n\n For Digit specifically, the company’s announced customer deployments, operating-hour disclosures and v5 plans provide a baseline against which new claims can be tested. If a new demo is described as a capability of v4, it should be compared with the v4 payload, runtime, end-effector and safety context. If it concerns v5, readers should distinguish a planned specification from an independently observed field result.\n\n For the wider humanoid sector, the same checklist applies:\n\n - Is the task valuable enough to justify a complex body?\n- Does the robot work at a rate that matters to the operator?\n- Can it recognise and contain failure?\n- What fraction of time requires a human?\n- What happens when the environment changes?\n- Are safety controls verified for the intended setting?\n- Are claims based on a fleet, or on a carefully selected run?\n\n The conference-floor Digit demonstration is worth covering because it shows the industry attacking a real bottleneck: the connection between getting to a task and physically completing it. It should be reported as an early signal of integrated capability, not as proof of general autonomy.\n\n That distinction is the practical advice for anyone evaluating humanoid robots now. Ask for the task definition, the denominator, the intervention policy and the recovery numbers. A robot that can finish a difficult motion once may be an impressive research system. A robot that can finish useful work repeatedly, explain when it cannot, and remain safe while people share its space is a deployable system. The distance between those two descriptions is where the robotics business will be decided.","available_translations":[{"language":"ar","title":"عرض Digit في IROS 2026 يوضح أن الاختبار المقبل للروبوتات الشبيهة بالبشر هو الاعتمادية القابلة للقياس","html_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=ar","markdown_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=ar","json_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=ar","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=ar"},{"language":"de","title":"Die Digit-Demo auf der IROS 2026 zeigt: Der nächste Humanoidentest heißt messbare Zuverlässigkeit","html_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=de","markdown_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=de","json_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=de","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=de"},{"language":"en","title":"IROS 2026’s Digit Demo Shows the Next Humanoid Test Is Measurable Reliability","html_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=en","markdown_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=en","json_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=en","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=en"},{"language":"es","title":"La demostración de Digit en IROS 2026 muestra que la próxima prueba humanoide será la fiabilidad medible","html_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=es","markdown_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=es","json_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=es","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=es"},{"language":"fr","title":"La démonstration de Digit à l’IROS 2026 montre que le prochain test des humanoïdes sera une fiabilité mesurable","html_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=fr","markdown_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=fr","json_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=fr","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=fr"},{"language":"pl","title":"Demonstracja Digita na IROS 2026 pokazuje, że kolejny test humanoida to mierzalna niezawodność","html_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=pl","markdown_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=pl","json_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=pl","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=pl"},{"language":"ru","title":"Демонстрация Digit на IROS 2026 показывает: следующий тест гуманоидов — измеримая надёжность","html_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=ru","markdown_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=ru","json_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=ru","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=ru"},{"language":"zh","title":"IROS 2026 的 Digit 演示表明：下一项人形机器人测试是可量化的可靠性","html_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=zh","markdown_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=zh","json_url":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=zh","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=zh"}],"_links":{"self":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=en","api":"https://publicasta.com/api/public/v1/channels/robots/articles/iros_2026_digit_mobile_manipulation_evidence?lang=en","html":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=en","canonical":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence?lang=en","markdown":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.md?lang=en","json":"https://publicasta.com/robots/iros_2026_digit_mobile_manipulation_evidence.json?lang=en","channel":"https://publicasta.com/api/public/v1/channels/robots","channel_articles":"https://publicasta.com/api/public/v1/channels/robots/articles","search":"https://publicasta.com/api/public/v1/search","documentation":"https://publicasta.com/api-docs#reading-publicasta","openapi":"https://publicasta.com/api-docs/openapi.json","llms":"https://publicasta.com/llms.txt"}}