SimPilot: prove the warehouse robot pilot in simulation before you buy the hardware

The problem

A single humanoid unit costs $50,000-$150,000 (Digit $100K-$150K, Apollo $50K-$100K), vendors will not put five-year service and parts pricing in writing before a pilot, and Indian SMEs cannot afford a failed automation bet. Logistics leaders need measured pick rates on their own SKUs and layouts before signing CapEx — but today only hardware vendors can credibly run pilots, and vendors run pilots to sell robots.

The idea

Why now

NVIDIA's Oct 21, 2026 drop put a production-grade robotics sim stack in the open for the first time: Isaac Lab 3.0 GA, the Isaac OS open runtime, openly available Cosmos 3.5 world models, and the Agile ONE SN reference humanoid. In the same month, GXO/Apptronik and Amazon/Digit pilots proved enterprises are buying pilots against hard KPIs right now, and DexForce's ~$147M raise validated sim-first training economics. The open stack removes the one thing that kept independent pilot services from existing: credible simulation tooling that isn't owned by a hardware vendor.

What it combines

Cosmos 3.5 (open world models + Data Factory for synthetic data) + Isaac Lab 3.0 GA (open RL training framework, multi-GPU) + Polibrain (predictive world model for robots) + Agile ONE SN (open reference humanoid). Any one alone is a tool; together they are a complete vendor-neutral pilot stack — simulation, synthetic data, policy training, and a reference robot — that previously only existed inside hardware vendors' labs.

MVP

Weekend MVP: take one public warehouse layout, build an Isaac Sim scene for a single tote-picking task, train/evaluate a pick policy, and generate a one-page pilot report (pick rate, cycle time, sim-to-real confidence interval, go/no-go). Deliberately skip: real robot integration, multi-site support, custom hardware drivers, and the white-label channel.

How it makes money

The moat

The largest incumbents are the robot vendors themselves — Agility, Apptronik, Figure, Tesla monetize unit sales. A vendor-neutral pilot service whose deliverable can be 'do not buy robots for this workflow' directly conflicts with their sales pipeline; they cannot offer it without undermining their own deals. NVIDIA sells GPUs and tools, not services. Systems integrators sell implementation hours, not 'don't implement' verdicts. The moat is channel conflict, not technology.

Distribution

Direct B2B to 3PL and warehouse ops heads in India (Delhivery, Ecom Express, Flipkart supply-chain, Reliance Retail logistics). Wedge: fixed-fee pilot reports per workflow (tote picking, palletizing, sortation), then annual 'sim validation' retainers as layouts change. Second channel: white-label the report engine to Indian systems integrators as a pre-sales de-risking tool.

Why it wins

Robot vendors (Agility, Apptronik, Figure) run pilots to close hardware deals; systems integrators sell integration hours. Nobody sells the independent verdict — especially 'don't buy.' AGILOX's end-to-end digital twin serves its own AMRs; DexForce is vertically integrated (its own sim software plus its own humanoids). A vendor-neutral service that can recommend against purchase is a different business, not a feature.

Risks

The sim-to-real gap: simulated pick rates may not transfer to a chaotic Indian warehouse floor, which would destroy the report's credibility. De-risk by starting with narrow, stable-geometry tasks (the 'narrow repeatable work' framing logistics leaders already accept), publishing the confidence methodology openly, and guaranteeing the first pilots against measured on-site numbers.

Build it with

• NVIDIA Cosmos 3.5 world models: now openly availableOpen world models plus the Cosmos Data Factory for generating synthetic training and evaluation data for the pilot scenes.
• NVIDIA Isaac Lab 3.0 GA + Isaac OS open runtime + Agile ONE SN humanoidThe open training framework and runtime the pilot simulations and policy evaluations run on.
• Polibrain (千诀科技): predictive world model for robotsPredictive world model to sharpen sim dynamics for contact-rich picking tasks.
• Natura-1: 50B robot foundation model trained on 1M hours of human videoPretrained manipulation prior to bootstrap pick policies before sim fine-tuning.

Repo to start from

simpilot: open pilot-report templates plus Isaac Lab scene configs and KPI measurement harnesses for standard warehouse tasks (tote picking, palletizing, sortation)

Evidence

• NVIDIA Advances Isaac Platform for General-Purpose Robotics (GlobeNewswire)
• Apptronik Apollo deployed at GXO warehouse (RobotWale)
• 5-year Cost of Humanoids in Warehousing (Logistics Business)
• Humanoid Robots vs. Indian Garment Workers: True Cost Comparison (Fractal Apps)
• DexForce trains humanoid robots in simulation (Glala Studios)
• Robot Fleet Traffic Management Software Market Report (Dataintelo)
• Robot as a Service Cost vs Owning a Tending Cell (ManufacturingMag)