Remy AI Wants to Teach Robots How Your Hands Actually Work

A camera can watch a warehouse worker lift a candle out of a bin and set it into a box. What the camera cannot see is the thing that matters most: how hard she squeezed. Press too soft and it slips. Press too hard and the glass jar cracks. That gap - between what a video records and what a hand actually does - is the whole business at Remy AI.

Remy AI is a San Francisco startup in Y Combinator's Winter 2026 batch, run so far by two people and one two-armed robot. The pitch is narrow and specific: dexterous, AI-powered robots for e-commerce third-party logistics warehouses, the 3PLs that pack and ship orders for mid-market retailers and Shopify brands. The bigger idea underneath it is about data - the kind nobody has been collecting.

01The problem, in newtons

Most of the warehouse world never got automated. Stacker cranes, shuttle systems, and fleets of autonomous mobile robots exist, but they only pencil out above a certain scale and predictability. A large distribution center with steady volume can justify a six-figure machine bolted to one task. A smaller 3PL cannot. Its inventory mix changes, its contracts are short, and its volumes swing with whatever its clients are selling that month. Rigid automation asks the warehouse to hold still. Warehouses do not hold still.

Remy's read is that flexibility, not raw strength, is the missing piece - and flexibility comes from data about how humans handle physical objects. That is harder to gather than it sounds.

There is far more training data online for cooking tutorials than for industrial work like wire harnessing, assembly, or suturing.Remy AI's founding observation

So the company collects it directly. Workers wear head-mounted cameras paired with tactile sensors that record three streams at once: vision, proprioception (where the body is in space), and contact force. The force channel is the unusual one. It captures the difference between a light grip on something fragile and a firm grip on something heavy - roughly 0.3 newtons versus 4 newtons - a distinction a camera-only system simply cannot register. The pipeline anonymizes and encrypts the footage.

02What the robot actually does

The product is a bi-manual mobile platform - two arms on a base that moves around the floor rather than sitting in a cage. It is built to cover several jobs a 3PL does every day: mobile picking, packing stations, kitting, and inbound receiving. The selling point is that it adapts to a new SKU without being reprogrammed, and slots into workstations the warehouse already has instead of demanding a rebuilt floor.

Under the hood are large pre-trained ML and robotics models fine-tuned per customer. Remy describes deployment as a short pipeline: photograph the customer's environment, simulate it, fine-tune the robot to that specific site, then harden it with real-world data and domain randomization. The company says this gets to production-grade reliability in days, not months - the stage where most robotics deployments quietly die.

03Who pays, and how

Remy sells through a robot-as-a-service model. Instead of a $100,000-$150,000 capital purchase tied to one task, a 3PL pays a monthly fee. That reframes the decision from a board-level capital expense into an operating line item, which is the only way the long tail of smaller operators can realistically say yes. The company is targeting roughly half the cost of traditional fixed automation, and it gets there by replacing mechanical complexity with software.

Replace mechanical complexity with AI intelligence.Remy AI

The customer profile is deliberately unglamorous: 3PLs handling fulfillment for Shopify-era brands and mid-market retailers - the roughly 70,000 of them in the US, inside a $1.2 trillion global logistics market growing about 8% a year, still run mostly by hand.

04Why this team

The two founders split cleanly across the problem. Oscar Brisset, CEO, is a former AI engineer at BCG's BCGX who built ML infrastructure for large enterprises. Ben Kaye, CTO, is an Oxford ML PhD with a CVPR 2025 Highlight paper on 3D reconstruction and three prior years writing firmware for life-critical medical devices at OrganOx. One end of the company knows how to move data and models at scale; the other knows how to make embedded hardware behave when failure is not an option.

That combination is the argument for defensibility. Remy is betting that three things compound: 3PL-specific dexterity data that only accumulates with deployments, a repeatable fast-deployment pipeline, and ML research depth in-house. None of the three is a moat alone. Together they are the case.

05Where it sits in the market

On one side are the incumbents of warehouse automation - stacker cranes, 2D/3D shuttles, AMR fleets - excellent at scale and predictability, poor at variety and small footprints. On the other side is the loud, well-funded race toward general-purpose and humanoid robots, most of it competing over the same publicly available video. Remy is threading between them: not a general humanoid, not a fixed conveyor, but a task-flexible robot for a specific, underserved buyer, trained on data it collects rather than scrapes.

Whether the dexterity-data bet pays off is still an open question. The company is early, small, and selling into an industry that has watched a lot of robotics promises arrive and leave. But the framing is unusually concrete for a seed-stage robotics pitch: a defined customer, a measurable cost target, and a modality - contact force - that most of the field has left on the table.

---

FAQQuick answers

---

Find Remy AI