Own the physical data plane.

Build the retrofit. Constrain its job.

The winning company is not “another dataset vendor.” It is the neutral layer that makes heterogeneous physical runs trustworthy, comparable, trainable, simulatable, and evaluable.

The retrofit should connect existing robots, teleoperation devices, and external sensors; establish a common clock and coordinate system; preserve raw evidence; emit model-ready views; create calibrated simulator counterparts; and close the loop from failure to targeted recollection. Commanded teleoperation belongs in a separate, hardware-specific safety envelope—not in the universal core. [E‑007]

“Robot data” is seven different goods.

A buyer does not purchase bytes in the abstract. Different data planes teach different things, carry different scarcity, and support different business models.

The full taxonomy also separates simulation ingredients from generated rollouts and evaluation evidence. This matters because simulation throughput is increasingly abundant while task authoring, contact fidelity, system identification, scenario validity, and real-world correlation remain difficult. [E‑001]

Real and synthetic data are complements

NVIDIA reported generating 780,000 synthetic trajectories—described as 6,500 human-demonstration hours—in 11 hours, and reported a 40% performance improvement from mixing synthetic and real data relative to real-only training. The correct conclusion is not that real data disappears: scarce demonstrations, target distributions, system identification, and real holdouts determine whether synthetic scale is useful. [S‑01]

The minimum valuable episode is causal

A useful episode contains task intent and success predicates; embodiment, firmware and controller state; calibrated observations; native robot state and requested/executed actions; outcome and intervention; clock and transform uncertainty; provenance, rights and lineage. LeRobot and RLDS provide important interchange structures, but the commercial evidence contract must also preserve calibration, contact, action semantics, privacy, and evaluation exclusion. [S‑02] [S‑03]

Five budgets. No single winner.

The buyer’s shortlist changes with the job: outsource collection, buy a rig, manage logs, simulate worlds, or keep building internally. The company must position against each substitution set.

Tier 1A — managed physical-data programs

Tier 1B — capture and teleoperation

Tier 2A — data and robot operations

Tier 2B — simulation and validation

The open seam is evidence continuity.

Managed providers operate programs. Hardware vendors provide rigs. Data platforms organize logs. Simulators create worlds. The company should own the integrity contract across all four.

The data-rights posture should be conservative: the customer owns its raw and task-derived data; the company retains adapter code, generic schemas, de-identified calibration statistics, aggregate quality patterns, generic residual models, and benchmark methods. Cross-customer trajectory reuse should be explicit opt-in. This creates trust without giving away the compounding technical system. [E‑009]

The Physical Data Plane

A rugged edge appliance is the distribution mechanism. The actual product is a certified chain from a physical run to a model result.

FIG. 02 — The product boundary. The universal core owns evidence integrity; hardware-specific control remains behind an entitled safety gateway.

Capture broadly. Control selectively.

The safe product expansion is a ladder. Each rung adds value—and a different class of engineering, safety and liability.

Interface availability is not permission to bypass safety. Real deployments must preserve vendor controllers, safety PLCs, interlocks, e-stops, watchdogs, workspace and force limits, and site risk assessment. [E‑008]

Sell a TaskPack, not a terabyte.

The commercial unit should describe a measurable capability. Recording hours and trajectories remain cost metrics inside the factory.

Positioning

For physical-AI teams that already own robots but cannot turn heterogeneous runs into reliable capability fast enough, the company is the hardware-neutral physical data plane and capability foundry. It retrofits the installed fleet, certifies every observation-action episode, expands scarce interactions in calibrated simulation, and proves model improvement on hidden tests—without replacing the customer’s robots, models, cloud or simulator.

When the company wins

• The customer has several robots or embodiments and repeats integration work.
• Temporal integrity, calibration, contact and outcome semantics affect policy quality.
• Simulation exists but is not demonstrably correlated with the real task.
• Failures and interventions are abundant but not converted into training and evaluation slices.
• Neutrality, customer data ownership and private evaluation matter.

When it loses

• A small lab has one stable robot and can operate an open Trossen/LeRobot workflow itself.
• The buyer primarily needs commodity annotation or thousands of operators immediately.
• The only pain is log visualization, fleet uptime or orchestration.
• The customer is fully standardized on a single vertical stack and already has excellent physical-data operations.

Battlecard shorthand

Twelve months to prove compounding.

The objective is not broad compatibility on paper. It is repeated deployment with falling marginal integration cost and demonstrated model lift.

Operating metrics

Design the company to be killable.

The retrofit thesis is attractive, but it can collapse into bespoke services, unsafe control work, or an appliance customers refuse to install. Six design partners should produce decisive evidence.

Main risks and mitigations

Six-part kill test

1. Median new-arm L1 integration still takes more than four engineer-weeks after three reference adapters.
2. Fewer than three of six partners will install the node on more than five robots.
3. Customers value annotation and search but will not pay for calibration and outcome instrumentation.
4. The system cannot find quality defects that materially affect policy performance.
5. Real↔sim calibration fails to improve scenario ranking or transfer on two task families.
6. Security and procurement consistently reject the appliance while accepting a pure software agent.

If the appliance fails but the evidence layer wins, pivot to an adapter SDK and managed calibration/evaluation service running on customer-owned edge hardware.