Better evidence for
robot intelligence.

Grounded in the work people need done

Our approach starts with industry workflows, with warehouses as the initial focus. We talk with the people who understand the work, identify a bounded task, and turn it into a repeatable physical evaluation with explicit success criteria.

We work with humanoid and manipulation model developers: comparing policies, testing robustness, and following performance across successive releases. That evidence can inform technical conversations with robot manufacturers, with the tested embodiment and its limitations kept in view.

Credibility comes before a ranking

We want a result to be useful because someone can inspect how it was produced. Protocols, trial accounting, failures, interventions, and uncertainty all belong beside the score.

Our first public benchmark is in development. We intend to make its methods, results, and supporting evidence open and verifiable. Until those results are ready, this site shares our approach and practical resources for designing evaluations.

Research and physical evaluation

We’re running benchmarks across robot embodiments and working with researchers at Penn, Illinois, MIT, the University of Michigan, UT Austin, and Georgia Tech. Our focus is the methods and physical execution needed to make model evaluations repeatable.

Coop is based in San Francisco, backed by a16z Speedrun, and a member of the NVIDIA Inception Program.

Work with the evidence

Model developers can explore our evaluation scopes. Researchers can use the protocol and trial-record templates. If you run an industry workflow, we welcome conversations about the tasks and operating conditions a useful benchmark should represent.