Capability review
Pressure-test the task, metric, current alternative, data evidence, and next milestone.
01 / Positioning
Sell a measured capability decision—not data volume.
The winning category is a Physical Capability Foundry: a neutral layer that links a physical failure, synchronized real evidence, calibrated simulation, and a hidden evaluation. WD-LANDSCAPEWD-COMP
Customer problem
A commercially important robot task performs in curated demos but misses an operational threshold. The team can collect more runs, yet cannot prove which data slice will change the deployment result.
Positioning statement
World Data turns one deployment failure into a validated acquisition plan—across the robots, models, cloud, and simulators the customer already uses.
First wedge: contact-rich manipulation
Industrial bimanual assembly, kitting, packing, depalletizing, or recovery offers measurable operational value, accessible validation, reusable task structure, and urgent data gaps. This is the recommended opening hypothesis—not yet a proven market decision. WD-LANDSCAPEINFERENCE
Do not start with
- commodity labeling or generic video collection;
- a marketplace spanning every physical domain;
- low-cost educational recording hardware;
- custom robot integration without a reusable evidence artifact.
02 / Qualification
A good design partner must pass six gates.
A16z recommends evaluating design partners on representativeness, urgency, and capacity. For physical AI, add access to the learning loop, proof reuse, and a real budget owner. A16Z-DPINFERENCE
Urgency and deadline20 pts
Representative task20 pts
Robot and site access20 pts
Model and evaluation access20 pts
Reusable proof rights10 pts
Budget and champion10 pts
A-fit: 75+ points, with no zero in urgency, access, or model/evaluation. A famous logo that cannot supply a task, champion, or weekly feedback is not a design partner.
≥7503 / Accounts
Prioritize teams that control both the robot and the learning loop.
The installed robot base is large, and leading model programs explicitly combine real, simulated, and Internet-scale data. The near-term buyer is the team accountable for turning that mixture into a deployment result. IFR-2025NVIDIASKILD
| Priority | Account type | Economic buyer | Buying trigger | First ask |
|---|---|---|---|---|
| P1 | Robot foundation-model labs | VP AI, Head of Data, Robotics Infrastructure | Skill plateau, new embodiment, or data mixture uncertainty | One frozen capability experiment |
| P1 | Robot OEMs entering customer pilots | VP Autonomy, Head of Learning, Deployment Engineering | Demo-ready task misses shift-level reliability | One pilot-blocking failure family |
| P2 | Manufacturers or 3PLs with internal physical-AI teams | Chief Automation Officer, Director of Robotics | Pilot has business value but cannot clear acceptance | Site, task, and release-gate evaluation |
| PARTNER | Robot, sensor, teleop, and simulation platforms | GM Platform, Ecosystem Lead, Strategic Partnerships | Platform needs a training-ready reference workflow | Adapter certification or co-sell reference |
| LATER | Operators without an internal learning team | Operations or automation leader | Generic automation need | Site or channel partnership—not core product sale |
Illustrative account universe—not customers or endorsements: Physical Intelligence, Skild AI, Figure, Agility Robotics, Apptronik, 1X, Boston Dynamics, BMW manufacturing, GXO, Amazon Robotics, Universal Robots, Franka, Kinova, Unitree, Trossen, and NVIDIA Isaac ecosystem teams.
04 / Offer ladder
Move from conversation to proof through four explicit products.
The ladder prevents a warm introduction from turning directly into an undefined pilot. Each stage has a decision, an artifact, and an exit. WD-LAUNCH
Capability Diagnostic
Freeze the bar, audit the record, capture a seed, calibrate a minimal twin, and run ablations.
Retrofit pilot
Retrofit two robots, produce accepted episodes and targeted simulation, then prove the acquisition result.
Capability loop
Failure clustering, targeted recollection, private release gates, additional robots, and adapter software.
Learn for free, without deploying.
Use five to ten trusted experts for problem interviews, prototype critique, pricing language, and introductions.
- No robot installation
- No custom integration
- No production deliverable
Deployment work is paid.
Once the work touches hardware, site access, calibration, simulation authoring, evaluation, or delivery commitments, it becomes a scoped Capability Diagnostic.
- Named task and metric
- Signed scope and rights
- Weekly technical cadence
05 / Partnership
Design the relationship around reciprocal evidence.
Recurring feedback, the real buyer and user, a defined term, and explicit expectations make the partnership useful. A16z recommends capturing these expectations in a contract or equivalent agreement. A16Z-DP
The partner contributes
- one operationally important task;
- robot and environment access;
- a technical champion and budget owner;
- a frozen model or hosted inference endpoint;
- weekly feedback and rapid decisions;
- permission to publish a sanitized method or explicit refusal.
World Data contributes
- a bounded six-week experiment;
- capture and calibration protocol;
- a small, high-information real seed;
- a minimal calibrated simulator counterpart;
- real-only, sim-only, mixed, and targeted ablations;
- a go/no-go recommendation for the 90-day pilot.
ScopeOne task, one robot family, one operational metric.
New embodiments or task families require a change order.
CadenceWeekly technical review with buyer and user represented.
Missing two consecutive reviews pauses the schedule.
AcceptanceArtifacts and experimental validity—not promised model lift.
Negative results are valid outputs when the experiment is sound.
SafetyRead-only capture by default.
Commanded teleoperation requires a separate interface and safety review.
RightsCustomer owns raw and derived task data.
Generic adapters and non-customer methods remain reusable.
ConversionDiagnostic ends with a 90-day pilot decision.
No indefinite “design partner” status.
06 / Diagnostic
The six-week engagement must end in a decision.
Each week resolves a different source of uncertainty. The output is a validated acquisition plan, not a pile of trajectories. WD-LAUNCH
Freeze
Baseline model, operational threshold, hidden test, meaningful lift, and leakage rules.
Exit: signed experiment charterAudit
Robot adapter, clocks, calibration, schema, rights, failure families, and coverage.
Exit: evidence gap mapAcquire
Success, failure, intervention, recovery, and a minimal calibrated simulator twin.
Exit: accepted seed TaskPackTest
Real-only, sim-only, mixed, and targeted subset ablations against the frozen test.
Exit: comparative resultDecide
Capability report, uncertainty, next failure slices, pilot economics, and remaining blockers.
Exit: go / revise / stopCommission
Scope two robots, one teleop path, volume, target TaskPack, data boundary, and evaluation.
Exit: 90-day pilot SOW07 / Pipeline
30scored accounts
12warm conversations
6capability reviews
3technical scopes
3technical scopes
2diagnostic proposals
1–2signed diagnostics
1reference pilot
Do not maximize inbound volume. Maximize the percentage of conversations that reveal a named failure, a measurable bar, a credible champion, and a path to hardware/model access.
08 / Outreach
Ask for a task, not “feedback.”
Every message should reference a known robot, deployment, task, or failure. The call earns the next experiment; it is not a generic company pitch.
Warm introduction request
I’m building World Data, a neutral evidence layer for robot learning. We start with one task below its deployment bar, retrofit the existing hardware, and test whether targeted real and simulated data improves a frozen evaluation.
[Name]’s work on [robot/task] looks unusually relevant. Would you be comfortable introducing us for a 45-minute capability review? I am not asking them to become a customer on the first call—the goal is to determine whether [specific failure] is a valid design-partner experiment.
Founder-to-buyer note
You are already operating [robot/system] against [task or milestone]. I suspect the blocker is not raw collection volume but knowing which failure slices are worth acquiring and whether simulation ranks them correctly.
World Data runs a six-week Capability Diagnostic around one frozen task: instrument the record, capture a high-information seed, build the minimal calibrated twin, and compare real-only, sim-only, mixed, and targeted data on a hidden test.
Would you spend 45 minutes deciding whether [specific task] is a credible experiment?
Disqualifying follow-up
The task is interesting, but it does not yet have the access and decision conditions required for a useful design partnership. The next step is to identify:
1. one operational metric;
2. one technical champion;
3. robot/environment access;
4. a frozen model or inference endpoint; and
5. the milestone this experiment would unblock.
When those are in place, we can rescore it for a Capability Diagnostic.
45-minute capability review
- 5 min: What deployment milestone is at risk?
- 8 min: Which task and failure family dominate?
- 7 min: What is collected today, and what is silently missing?
- 7 min: What model and hidden test can be frozen?
- 6 min: Which robots, operators, sites, and simulators are accessible?
- 5 min: What data, CAD, SOP, or process must remain inside the customer boundary?
- 4 min: Who owns budget and implementation?
- 3 min: Decide: diagnostic scope, prerequisite work, or no fit.
09 / Economics
Six-week diagnostic$50k–$100k
Target range for a bounded capability experiment. Credit a negotiated portion toward the 90-day pilot when conversion happens quickly.
90-day retrofit pilot$150k–$350kExcludes unusual hardware, site compliance, high-risk control access, and customer-specific exclusivity.
Raw robot and sensor evidenceCustomer-owned
Derived task datasetsCustomer-owned
Customer scenes, CAD, SOPs, process IPCustomer-owned
Generic adapter code and schemasWorld Data
De-identified calibration methodsContracted
Cross-customer dataset useExplicit opt-in
Sanitized methodology or resultWritten approval
Commercial rule: never trade all reusable learning for a prestigious logo. If exclusivity blocks adapter reuse, calibration methods, sanitized methodology, and proof, the economics must compensate for the lost company asset. Contract terms require qualified legal counsel; this strategy is not legal advice.
10 / 90-day plan
Build the company in three founder sprints.
The objective is not a public launch. It is a repeatable offer, one credible TaskPack, two paid diagnostics, and the conditions for a reference pilot.
Interview and score
- Build a 30-account list from existing contacts.
- Run ten private page reviews and six capability interviews.
- Score every task with the 100-point partner rubric.
- Freeze the diagnostic charter, data-rights default, and sample TaskPack.
- Choose one task family and two robot targets.
Exit: three A-fit scopes and two buyers willing to discuss a paid diagnostic.
Scope and contract
- Deliver two diagnostic proposals.
- Complete security, safety, site-access, and rights diligence.
- Establish baseline/evaluation requirements.
- Sign one or two paid diagnostics.
- Document every objection and competitive alternative.
Exit: signed scope, champion, budget owner, model endpoint, robot access, and weekly cadence.
Run the first slice
- Complete the first audit and accepted seed.
- Produce the minimal real↔sim correlation artifact.
- Run at least one data-mixture ablation.
- Draft the 90-day pilot SOW.
- Secure permission for one sanitized result or methodology note.
Exit: a defensible go/revise/stop decision and a reusable reference artifact.
11 / Metrics
Measure learning velocity and commercial proof together.
Traffic is secondary. In a low-volume, high-value founder motion, task quality, technical access, deal progression, and reusable evidence determine whether the company is compounding.
Qualified capability reviewsDo conversations reveal a named task, metric, and deadline?
Review → technical scopeDoes the buyer grant access to the model and robot boundary?
Scope → paid diagnosticWill the budget owner pay for evidence, not engineering hours?
Time to first accepted episodeIs the retrofit becoming reusable?
Calibration-valid session rateCan silent data poison be detected?
Failure → targeted data cycleIs the loop getting faster?
Reusable artifact ratioWhat survived beyond one customer?
Kill or pivot gates after six design partners
- new read-only arm integration still takes more than four engineer-weeks;
- fewer than three partners will deploy the node on more than five robots;
- buyers value search or annotation but not calibration and outcome evidence;
- the system cannot detect defects that materially change policy performance;
- real↔sim calibration does not improve ranking or transfer on two task families;
- procurement rejects the appliance but accepts software-only agents.
If appliance deployment fails while evidence software is valued, pivot toward an adapter SDK plus managed calibration and evaluation running on customer hardware. WD-COMP
12 / Sources
Evidence and assumption ledger.
The published strategy distinguishes verified source claims, World Data inferences, and commercial hypotheses that require customer validation.
SOURCE directly groundedINFERENCE strategic synthesisHYPOTHESIS target to validate
WD-PRODUCTWorld Data product brief: users, purpose, positioning, operating context, and claim boundaries. Internal source, 2026-07-28.
WD-LAUNCHRobotics Intelligence Landing Page and Network-Led Launch Plan: offer ladder, ICP, warm-launch sequence, funnel, proof, and planning assumptions. Internal source, 2026-07-28.
WD-COMPPhysical-AI Data Competitive Products and Retrofit Strategy: competitive alternatives, retrofit wedge, economics hypotheses, moat, risks, and kill criteria. Internal source, 2026-07-28.
WD-LANDSCAPEThe State of Physical-AI Data: buyer segments, build/buy boundary, capability-foundry thesis, and task-wedge analysis. Internal source, 2026-07-28.
A16Z-DPSeema Amble and Jennifer Li, “A Framework for Finding a Design Partner”, 2022. Supports representativeness, urgency, capacity, warm introductions, recurring feedback, and defined terms.
IFR-2025International Federation of Robotics, “Global Robot Demand in Factories Doubles Over 10 Years.” Supports installed-base context.
NVIDIANVIDIA Humanoid Robots. Supports the mix of real captured, synthetic, and Internet-scale training data.
SKILDSkild AI, “Building the General-Purpose Robotic Brain.” Supports the role of simulation, Internet video, and targeted real data.
BMWBMW Group, Physical AI program and Leipzig deployment, 2026. Demonstrates operator-side physical-AI teams and real production pilots.