A robot picks something up. On video, it looks like progress. Put the same robot on a warehouse shift, however, and the questions multiply. How many items an hour? How often does it stop? Who resets it? The interesting moment arrives when the camera keeps rolling after the impressive bit. That is where Orion VC has placed one of its latest bets.
- Orion backs AI companies at pre-seed and seed, with an emphasis on technical founders.
- Its first investment, Feedback Intelligence, was acquired by ActiveCampaign.
- Its September 2026 Positronic investment backs real-world testing of robot AI.
In September 2026, the San Francisco investor joined Positronic Robotics’ $2 million pre-seed round, led by 33East. Positronic tests AI models on physical robots and measures throughput, reliability and failure. The round’s headline number belongs to the whole syndicate. The more revealing detail belongs to the experiment: at its public benchmark’s launch, the best model achieved roughly five percent of human throughput.
The five-percent problem
Five percent is an awkward number for a promotional film. For someone deciding whether to deploy a robot, it is useful. Positronic’s PhAIL benchmark begins with bin-to-bin order picking. It compares models with human and teleoperated baselines, and publishes complete episodes, including failures and safety stops. The unit of interest is productive work over time. A successful grab tells only part of that story.
Best model’s throughput
relative to the human baseline
There is a business inside this unglamorous distinction. Positronic offers a public leaderboard for open models, while its core business is private evaluation during training. Orion names Runway as a design partner. The new capital is intended to expand the robots, tasks and models being tested. The public test makes the problem visible; private testing helps laboratories work on it before the next release.
Positronic’s founders met at Google, where evaluation of Search was an established engineering discipline. Their observation was that physical AI lacked an equivalent. It is a wonderfully practical origin: take a capability that one field treats as ordinary and build it for another field that urgently needs it. Orion’s investment is a bet that independent measurement becomes more valuable as the surrounding market matures.
The first bet listened for friction
The connection to Orion’s first investment is striking. Feedback Intelligence studied conversations between people and AI agents, turning unstructured exchanges into information about satisfaction, intent and friction. A conversation counter can show activity. It cannot, by itself, explain why a person left disappointed. Feedback Intelligence worked on that second question.
ActiveCampaign announced its acquisition of the company in February 2026. Founder Chinar Movsisyan and her team joined its technology and product organization. The buyer described evaluation infrastructure that would help its Active Intelligence system improve through actual interactions. For Orion, the transaction supplies a concrete early outcome. For the reader, it makes the investment thesis easier to see: AI products need ways to discover where their apparent competence runs out.
“Independent measurement becomes infrastructure as a field matures.”
Sergey Dean · September 2026
The two companies measure different things. One examines conversations; the other watches machines attempt physical work. Their common territory is the distance between a system’s promise and a customer’s experience. That is an attractive place to look for a business because the customer already has a reason to care. A laboratory needs feedback. An operator needs dependable output. Neither can pay its bills with a lovely demo.
Early enough to be uncomfortable
Orion’s January 2026 review described four portfolio companies, all backed while still in stealth. It said each had raised follow-on capital and that Orion had led two of those rounds. Those are the firm’s reported milestones, rather than a measure of investment returns. Still, they show the intended entry point: before broad attention gives everyone the same reassuring signals.
Its stated remit spans AI infrastructure, physical AI and vertical AI. The current website puts particular emphasis on infrastructure and physical systems, technical depth and immigrant founders building in the United States. That is a narrower proposition than simply investing in anything with an AI label. It asks whether the difficult part of a company can become an advantage that rivals struggle to reproduce.
The partners bring relevant experience to that question. Managing partner Sergey Dean’s background includes mathematics, McKinsey and leading AI at Mindbox. General partner Armen Fljyan previously worked at SmartGateVC, sourcing AI investments and supporting founders. Orion presents practical involvement as part of its offer, including help with hiring and the work of building a young business.
A useful room, before a useful round
Capital is one part of that offer; access is another. Orion’s events place founders near people who build, buy and fund technology. Its October 2026 SF Tech Week panel brought together the founders of Activeloop, Campfire, robot.com and Wirestock. The subject was the infrastructure AI still needs: data, compute and tooling. The guest list describes several different routes from a model to something a business can use.
That activity also locates Orion in the market. It is an early specialist investor, with technical founders on one side and limited partners supplying investment capital on the other. Its service is financing and founder support, rather than a subscription product. Other early AI funds and technically informed angels offer alternatives. Orion’s case rests on the usefulness of its judgment and relationships at a stage when neither revenue nor reputation can do all the persuading.
Keep the camera running
There is something here a founder can copy without raising a round. Choose a task a customer actually needs completed. Decide how success and failure will be recorded. Compare the result with the customer’s existing method. Keep the awkward attempts in the record. This is an inference from the portfolio, rather than an Orion instruction manual, but it is a more useful starting point than polishing a demonstration.
The conditions matter. A benchmark only helps if its tasks resemble the work someone will buy, and proprietary data only creates an advantage if it improves a product customers want. Deep engineering can consume money without producing demand. Orion’s distinctive wager is that these difficult, measurable problems deserve attention early. The appealing part is the insistence on looking closely. Leave the camera on. Count what happens next.