"We are obsessed with breaking & improving robot policies."
Munich robot-evaluation firm running a Franka FR3 Duo farm at the Deutsches Museum; its PAW-GEN-10 benchmark logged 3,600 real rollouts per policy.
Poke & Wiggle GmbH is a Munich company that evaluates robot foundation models and vision-language-action (VLA) policies on real hardware. It is registered at the Munich local court under HRB 301692, with Nicolas Keller and Jose Ramon Medina Hernandez listed as representatives, and states offices in Munich, San Francisco and London. Its robot operations are located at the Deutsches Museum in Munich, where it runs what it describes as the largest independent farm of Franka FR3 Duo bimanual robot stations.
Each benchmarking station pairs two Franka FR3 arms on a Franka Duo pedestal with Robotiq 2F-85 grippers, Intel RealSense D405 wrist cameras and a ZED Mini head camera. Operators in VR headsets teleoperate the stations to collect demonstration data, and the same stations then execute customer or open-source policies under controlled, repeatable conditions.
The company sells three services. Reality Check benchmarks a robot policy on standardised tasks with real robots, lets the customer compare it with other models and inspect the evidence behind each score, and offers private results or publication on the public Reality Check Leaderboard. Loop gives AI labs remote access to the robot farm to run successive checkpoints across real and simulated stations, with results in a customer data portal; Poke & Wiggle also helps debug training pipelines, collects targeted data and verifies improvements. Application Benchmarking, aimed at industrial teams, turns a customer's own task into a benchmark, tests candidate policies against it and reports what works, what fails and what still needs development.
On 29 September 2026 the company released PAW-GEN-10, a benchmark of ten real-world environments covering basic skills, tool use, high-precision assembly and bimanual handling, such as sorting screws, clamping an electrical component, connecting a DC jack and routing a cable through closed hoops. Five environments are held out and unpublished. Four open policies, MolmoAct 2, Pi 0.5, DiT-Flow and GR00T N1.7, were each fine-tuned on 10, 100 and 300 demonstrations per environment and evaluated over 3,600 rollouts per policy. MolmoAct 2 scored highest at 44% task success with 300 demonstrations, and in 86% of failed episodes the hardware was fine and the policy simply did not finish the task. Metrics include success and progress, execution time, motion smoothness, jerk, safe failure rate and maximum contact force.
In September 2026 world-model developer Odyssey named Poke & Wiggle as a collaborator for assessing how its Odyssey-3 model adapts across robot types and control systems. The company also runs a public Discord community.
Primary type & automation activities this supplier delivers:
Real-robot benchmarking of robot policies on standardised tasks, private or on the public leaderboard.
Ten-environment real-world manipulation benchmark for VLA policies with reproducible scenes.
Remote access to a Franka FR3 Duo robot farm to iterate policy checkpoints on real and simulated stations.
Turns an industrial customer's task into a benchmark and tests candidate robot policies against it.
Contact Poke & Wiggle
WEBSITE
https://pokeandwiggle.comHEADQUARTERS
Germany
Company Facts
Founded
-
Primary Role
Software/Algorithm
Company Size
-
Primary Region
Europe
Annual Sales
-
Funding Stage
-
Funding Total
-
Keywords