Decentralised News · 2027 planning series · Sources reviewed 22 November 2026
What Matters
A robot demonstration shows a capability under particular conditions. A purchasing decision needs evidence about a complete shift: accepted work, downtime, human assistance and total cost. DN’s 2027 model turns those inputs into planned output and cost per accepted unit. It does not rank humanoids, certify workplace safety or convert a vendor’s target into a measured result.
The useful robot story begins after the applause
The most revealing detail in Figure’s account of its BMW deployment is not a dramatic movement. It is the work specification. The company described sheet-metal loading and identified cycle time, placement accuracy and human interventions as three operating measures. Its stated goals included greater than 99% placement success per shift and zero interventions. Goals are not the same as published achieved rates. Figure’s November 2025 deployment account makes that distinction possible because it names the task and the measures.
BMW’s own February 2026 account reports roughly 1,250 operating hours and more than 90,000 components moved during the Spartanburg pilot. These are customer-reported operational totals, not a DN audit of unit economics. They establish a more concrete starting point than an isolated clip, while leaving important purchasing questions unanswered.
A buyer still needs to know how much work met acceptance criteria, which support resources were required and whether the same result can be delivered in the buyer’s environment. The transition from demonstration to business case occurs in those missing denominators.
Evidence Block
Reviewed 22 November 2026 for 2027 planning. Sources include robot manufacturers, a customer deployment account and NIST’s measurement research. Commercial statements are identified as supplier or customer reports. DN has not run a physical robot trial, measured the named systems’ intervention rates or inspected their contracts. All calculator defaults below are hypothetical, unrelated to any named vendor and expressed in one user-selected currency.
1. Name the job before choosing the body
A humanoid shape is a design characteristic. It is not an acceptance criterion. Start with the required outcome: move a specified container between two locations, position a component within tolerance, inspect a defined surface or complete another bounded task. Record the object range, environment, quality requirements and deadline.
Then consider the alternatives. A fixed arm, conveyor, mobile platform, workplace redesign or human-assisted process may solve the same problem. The comparison should use the same accepted output and cost boundary. Selecting a humanoid because it looks adaptable can obscure whether adaptability is required for this particular job.
Separate present requirements from possible future tasks. A useful option to learn another job can have value, but the base purchase case should identify which skills are available and supported today. A planned capability belongs in a separate scenario with its own evidence and dependency.
For example, Unitree’s G1 product page states that some showcased functions remain in development and that parameters vary by configuration and scenario. The appropriate buying response is to verify the quoted configuration and supported task, not to assume every demonstrated movement is delivered as a production feature.
2. Distinguish control modes without treating assistance as failure
There are several ways to complete a task: a human can remotely control the motion, a system can execute a prepared routine, an operator can select high-level actions, or the robot can act within a defined autonomous policy. A deployment can combine these modes. A useful record states which mode applied to each completed task.
Human assistance is not inherently disqualifying. A remotely operated machine may be valuable because it changes where a person must work or which physical burden they face. Its economics should include that person. A product marketed as autonomous needs evidence for the autonomy claim at the relevant task boundary.
Ask who starts the task, handles unexpected objects, resets failed attempts, approves recovery and prepares the next work area. A robot can execute its central movement independently while relying on people before and after that movement. Those surrounding activities belong in the operating record.
Unitree’s official support material, for example, provides remote-control, calibration and maintenance documentation. The existence of those functions does not reveal how a particular video was produced. It does show why a buyer should ask for explicit operating-mode and support disclosures rather than infer them from appearance.
3. Use the full shift as the time boundary
Begin with scheduled hours. Then identify unavailable time: charging or battery changes, setup, maintenance, resets, waiting for materials and any other interval in which the cell cannot perform the defined work. Use mutually exclusive categories so that the same ten minutes is not subtracted twice.
Available operating time is scheduled time minus those unavailable intervals. This is a planning definition for the DN model, not a claim that every manufacturer reports availability identically. Record the definition beside the number whenever comparing products or sites.
Cycle time should then measure an attempted task while the cell is operating. If the reported cycle already includes average pauses, do not subtract the same pauses again as downtime. The source of many attractive business cases is an inconsistent boundary: a fast demonstration cycle is multiplied by every scheduled hour even though the real shift includes other work.
A machine can also be available while the wider process has no useful work for it. Distinguish technical availability from demand. If a site needs only 150 units, a scenario capable of 500 does not create 350 additional units of economic value automatically.
4. Count accepted output and retain rejected work
Define acceptance before the pilot. For a container move, it might include the correct object, correct destination, intact contents and completion within the required time. For positioning work, the record needs the agreed tolerance and inspection method. The criterion must describe the buyer’s usable output.
Count attempted tasks separately from accepted units. Repeated attempts on one item can inflate activity without increasing production. Give the business item a stable identifier and record retries, rejection, rework and final disposition against it.
A rejection still consumes machine time and may consume parts, energy and human effort. Keep those costs in the numerator. Excluding unsuccessful work from both cost and output makes the system appear more efficient than the process the buyer must operate.
Acceptance also needs a follow-up boundary. A component that looks correctly positioned but fails a later quality check should not remain a permanent success in the robot report. Choose an observation window appropriate to the workflow and explain how later corrections revise the cohort.
5. Measure the human work around the machine
Record hands-on assistance, remote operation, planned supervision, maintenance, setup and incident investigation. Several people working simultaneously create several labour-minutes, even though only one minute passes on the clock. Use role-specific loaded costs when they differ materially.
DN’s simple calculator uses intervention minutes per attempted cycle plus other allocated labour hours. It assumes those categories do not overlap. A more detailed deployment should retain event-level logs showing who did what, when and for which task.
Track queueing as well as labour. One technician might support several robots comfortably until three stop at once. Average assistance minutes do not demonstrate that every machine can resume within the required time. Correlated stoppages and the availability of specialist support belong in the deployment decision.
This extends DN’s human-rescue accounting framework into a physical work cell. The additional issues are machine availability, cycle time, material quality, installed hardware and the capacity of the surrounding production process.
DN Robot Workday Cost Model
The calculator uses a scheduled shift, unavailable hours, active cycle time, acceptance rate, intervention time, other labour and daily cost allocations. It calculates available hours, attempted cycles, expected accepted output, human hours and cost per accepted unit.
Planned attempts = the whole number of cycles that fit into available seconds. Expected accepted output = planned attempts × acceptance rate. Human hours = attempted cycles × intervention minutes ÷ 60 + other allocated labour hours. Total included cost = daily fixed allocation + energy + other non-labour cost + human hours × loaded hourly rate.
Cost per accepted unit divides that total by expected accepted output. The output is a scenario, not a measured forecast. Expected output can be fractional because it applies a rate to planned attempts; an observed production ledger should use actual whole accepted counts instead.
Where a comparable manual process is available, enter its accepted units per hour. The model calculates the labour cost of producing the same expected quantity at the same hourly rate. This narrow comparison excludes unentered manual overhead and does not establish cash savings or a complete investment return.
DN Robot Workday Cost Model
Plan one shift in one currency. Defaults are hypothetical. Keep cycle time, downtime and labour categories separate, and compare equivalent accepted work. No vendor performance is implied.
This instrument runs locally in your browser. It does not connect to accounts, fetch evidence or submit instructions. Use aggregate values and non-sensitive references. Downloaded records contain your inputs.
Machine-readable illustrative record
An illustrative record of the example assumptions, not an observed transaction or provider test. A null calculation timestamp identifies this static example.
{
"schema_version": "1.0.0",
"article_id": "N007",
"evidence_basis": "illustrative",
"source_review_date": "2026-11-22",
"computed_at": null,
"inputs": {
"shift": "8",
"downtime": "2",
"cycle": "120",
"acceptance": "95",
"intervention": "0.5",
"other_labour": "1",
"wage": "30",
"fixed": "50",
"energy": "10",
"other_cost": "20",
"manual_rate": "60"
},
"status": "Illustrative workday calculation",
"outputs": {
"methodology": "DN robot workday cost model v1.0",
"hypothetical": true,
"inputs": {
"shift": 8,
"downtime": 2,
"cycle": 120,
"acceptance": 95,
"intervention": 0.5,
"other_labour": 1,
"wage": 30,
"fixed": 50,
"energy": 10,
"other_cost": 20,
"manual_rate": 60
},
"available_hours": 6,
"availability_percent": 75,
"attempts": 180,
"expected_accepted": 171,
"human_hours": 2.5,
"total_cost": 155,
"cost_per_accepted": 0.9064327485380117,
"manual_direct_labour_cost": 85.5,
"excess_cost": 69.5,
"certifies_safety": false
}
}Worked example: the cheap cycle produces an expensive shift
Hypothetical example, not vendor performance. A robot cell is scheduled for eight hours and unavailable for two. An active cycle takes 120 seconds. The six available hours contain 180 attempted cycles. At an assumed 95% acceptance rate, planned accepted output is 171 units.
Each attempt needs an average of half a minute of human assistance, totalling 90 minutes. Another hour covers allocated setup and supervision. At 30 currency units per labour hour, the 2.5 human hours cost 75. Add a daily fixed allocation of 50, energy of 10 and other non-labour costs of 20. Total included cost is 155, or approximately 0.91 per accepted unit.
A hypothetical manual process producing 60 accepted units an hour would require 2.85 hours for the same quantity. At the same hourly rate, its direct labour cost is 85.50, or 0.50 per unit. Under this deliberately limited comparison, the robot scenario costs 69.50 more. The result changes if omitted benefits or costs are material, but the attractive active cycle alone does not establish an advantage.
Now change one assumption rather than all of them at once. Reducing intervention time to a tenth of a minute lowers assistance to 18 minutes and total labour to 1.3 hours. Included cost falls to 119, or roughly 0.70 per accepted unit. That improvement is meaningful while still falling short of the assumed manual labour benchmark.
The example exposes which input deserves measurement next. It does not imply that every robot needs this amount of help or that labour replacement is the only reason to deploy one.
6. Put the installed system in the price
The robot’s purchase price is one line. The operating system may also require tooling, fixtures, sensors, charging equipment, networking, software, integration, spare parts, transport, commissioning and training. Request an itemised quote for the delivered configuration and the task to be supported.
For an owned system, allocate the relevant installed cost over a stated useful life and operating schedule, with residual value separately justified. For a service arrangement, use the actual recurring and usage-based charges. Do not count both the full purchase price and its depreciation allocation in the same operating-period total.
Keep financing, tax and cash-flow analysis separate from the simple unit-cost model unless explicitly added. A low daily depreciation figure does not mean the business can fund the initial installation. A lease can reduce upfront cash while retaining minimum commitments that matter if the pilot disappoints.
Maintenance exclusions deserve particular attention. Identify who pays for damage, replacement components, software changes and support outside the standard window. The model’s “other cost” field is a place to allocate documented items, not a substitute for reading the agreement.
7. Pilot in the conditions the business actually faces
NIST’s robot agility programme focuses on measuring performance across tasks, unexpected events, variations and reconfiguration. That research direction supports a practical purchasing principle: test the work and the changes that matter to the intended environment. It is not a certification of the products discussed here.
DN proposes a pilot matrix with an ordinary task, a representative object variation, a planned changeover, an authorised recovery exercise and a sustained run. Use only trials approved by the site’s qualified safety personnel and relevant operating procedures. Do not improvise hazardous failure tests to create a compelling video.
| Pilot question | Record required | What it helps decide |
|---|---|---|
| Can it perform the defined job? | Attempts, acceptance criteria and inspected results | Basic task fit |
| What happens across the shift? | Scheduled time and exclusive downtime categories | Usable capacity |
| Who assists it? | Mode, role, minutes and reason for intervention | Staffing and labour cost |
| What changes between products? | Configuration, tooling and changeover time | Value of flexibility |
| What does it cost to keep working? | Installed cost, service scope and maintenance records | Operating commitment |
Retain all runs within the declared cohort. If the vendor excludes a failed day or changes the software midway, preserve both versions and explain the comparison. A selected highlight reel cannot supply the denominator required by a purchasing decision.
8. Commercial deployment is evidence, with limits
On 26 October 2026, GXO and Agility Robotics announced an expanded partnership for additional US sites in 2027. Their release reports more than 100,000 totes moved at the Flowery Branch facility and describes expected deployments beginning in early 2027. The completed activity and the future deployment plan should remain separate claims.
The announcement is evidence of a named commercial relationship and reported work. It does not publish the full customer cost, every intervention, all rejected moves or an independently audited comparison with alternatives. A buyer can treat deployment experience as relevant while still asking for the operational data needed for its own site.
Likewise, “contributed to vehicle production” does not mean a humanoid built an entire car. Preserve the task boundary when repeating a deployment claim. Expanding a narrow, useful contribution into general labour replacement makes the evidence less informative.
DN Alpha Thesis
The important commercial unit for physical AI may be the accepted workday: a defined quantity of useful output delivered within a shift, with the human support and installed system included. Buyers that measure this unit can compare very different machines without making the humanoid shape the investment thesis.
Better autonomy could improve that unit by reducing assistance, widening the range of supported tasks or shortening changeovers. Better hardware, fixtures and service could improve it too. The thesis would weaken if these measures failed to predict useful deployment outcomes or if a simpler task-specific metric consistently explained purchasing value better.
Methodology and what the model leaves out
DN separates manufacturer specifications, reported customer activity, operating targets and its own proposed arithmetic. No named robot inherits the calculator’s default inputs. The article’s public examples were checked against primary pages; their evidence dates are retained rather than presented as events occurring in 2027.
The deterministic model assumes one cell, one shift, one cycle time and stable input rates. It does not simulate queueing, learning, correlated failures, variation in object difficulty, financing, taxation, safety risk or downstream production losses. It also does not value ergonomic improvements or other benefits without a separate measurement basis.
Zero accepted output makes cost per accepted unit undefined, not free. Unavailable time cannot exceed the shift. Percentage inputs must remain between zero and 100. A negative cost or a zero cycle time is an invalid scenario. Where facts are unknown, keep them unknown and collect evidence instead of entering a favourable default.
For the distinct question of software agents interpreting images and acting on screens, see DN’s Multimodal Agent Readiness Test. This physical-workday edition contains no affiliate hardware recommendation, paid ranking or product safety certification.
Frequently asked questions
Does a robot demonstration prove that it can replace a worker?
No. A demonstration establishes a capability under stated conditions. A staffing or purchasing decision needs evidence across the complete workflow, including assistance, downtime, accepted output and installed cost.
Is teleoperation a failure?
No. Remote operation can be valuable. Its human effort must be included, and an autonomy claim must describe the actual task boundary and operating mode.
Are the calculator defaults measurements of Figure, Digit or Unitree?
No. Every default is hypothetical and unrelated to the named manufacturers. Public deployment statements are discussed separately and attributed to their sources.
Why can expected accepted output be fractional?
The planning model multiplies whole attempted cycles by an assumed acceptance rate. That produces an expected value. Actual production reports should use observed whole accepted counts.
What happens if there are no accepted units?
The model reports included cost and any human work, but cost per accepted unit is undefined. Zero output is not free production.
Does lower modelled unit cost equal cash savings?
No. The result depends on cost boundaries and whether capacity can be used. Reduced staff time does not automatically reduce payroll or justify an investment.
Does this worksheet certify a safe robot deployment?
No. It is a planning calculator. Workplace safety, integration and operating procedures require their own qualified assessment.
Primary sources and review record
- Figure: F.02 Contributed to the Production of 30,000 Cars at BMW. Vendor report, task and KPIs; placement/intervention values explicitly identified as targets, not measured results. Reviewed 2026-11-22.
- BMW: Humanoid robot production pilots, February 2026. Customer-reported 1,250 operating hours and 90,000+ moved components; bounded pilot task. Page displayed publication 27 February 2026 despite URL suffix. Reviewed 2026-11-22.
- Unitree G1 product page. Configuration/scenario caveats and some sample functions still in development; no current price used because search snippet and opened page differed. Reviewed 2026-11-22.
- Unitree Explore support and manuals. Remote control, calibration and maintenance support exists; no inference about any particular demo control mode. Reviewed 2026-11-22.
- NIST: Agility Performance of Robotic Systems. Measurement research on task changes, failures, variation and retasking; not product certification. Reviewed 2026-11-22.
- GXO and Agility expand partnership, 26 October 2026. Named commercial relationship, reported 100,000+ totes moved, expected 2027 new-site deployment; completed and forward-looking claims separate. Reviewed 2026-11-22.