Poke & Wiggle is launching Reality Check, a public leaderboard and benchmark that measures how well robot manipulation policies actually perform when tested on physical hardware. The first release, PAW-GEN-10, puts four open robot policies through ten real-world environments with 3,600 reproducible rollouts per policy, totaling 14,400 rollouts , and studies three axes of generalization: task-specific data regimes, spatial generalization to unseen object placements, and embodiment-specific pre or mid-training . The environments cover bimanual skills spanning basic manipulation like sorting and pouring screws, tool use, high-precision tasks such as connecting a DC jack and routing cable through hoops, and coordinated two-arm work .
The launch matters because much of the public discourse around physical AI still relies on cherry-picked demo videos rather than statistically meaningful evaluation. Poke & Wiggle argues that rollouts should be treated as measurements, which means running many of them and keeping conditions repeatable so nothing changes between runs except the policy under test . To support that, the team operates the largest independent Franka FR3 Duo farm, helping groups building frontier world models and robot foundation models evaluate them in the real world, uncover failure modes, and close the loop with high-quality data , with the physical operations housed inside Munich's Deutsches Museum.
The company was founded in 2025 by CEO Nicolas Keller and CTO José Ramón Medina Hernández, with offices in San Francisco, London, and Munich. Reality Check is aimed at robotics research groups, foundation model teams, and industrial buyers who need evidence-backed comparisons before committing to a policy or a deployment, and Poke & Wiggle plans to keep publishing rollouts and expanding the leaderboard as new models are added.
Comments (11)
Sign in to join the discussion.
R
Ravi Subramanian14d ago
14,400 rollouts is a spicy number to open with. What's the wall-clock cost of one full leaderboard run on the FR3 duo? Asking because my advisor will.
T
Tom Carver14d ago
Would love to chat about the model selection story before this hits the wider press cycle. DMs open, promise I'm nice.
L
Lena Klein14d ago
Finally a robotics eval that admits confidence intervals exist. Half the manipulation papers out there report one seed and call it science.
H
Hiroshi Tanabe14d ago
Real question: are the rollouts scripted resets or does a human reset the scene between trials? The variance story changes a lot depending.
T
tomiwa14d ago
That video opens strong but the middle third drags. Would've cut straight from the gripper fail to the leaderboard reveal, saved everyone 8 seconds.
M
Maxine Petrova14d ago
A 1/10 thread on launch day is a bold commitment. If tweet 7 isn't the money shot I'm rioting in your replies.
K
Kwabena O.14d ago
Cute leaderboard. Who pays for this and how often do they re-run it before the novelty wears off?
P
Priya Deshmukh14d ago
The object placement grid is the part everyone will steal and no one will cite. Screenshot it now.
D
dani f.14d ago
The thumbnail is doing zero work. A gripper mid-fail would've outperformed the clean hero shot by a mile.
O
Olumide A.14d ago
How big is the team behind this? Standing up 14k physical rollouts sounds like six people who haven't slept since October.
S
Sasha Volkov14d ago
Manual benchmark suites in 2025 feel quaint. Next version should be an agent that invents adversarial tasks on the fly and grades itself.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Poke & Wiggle
We are obsessed with breaking & improving robot policies.
We are obsessed with breaking & improving robot policies.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Poke & Wiggle is launching Reality Check, a public leaderboard and benchmark that measures how well robot manipulation policies actually perform when tested on physical hardware. The first release, PAW-GEN-10, puts four open robot policies through ten real-world environments with 3,600 reproducible rollouts per policy, totaling 14,400 rollouts , and studies three axes of generalization: task-specific data regimes, spatial generalization to unseen object placements, and embodiment-specific pre or mid-training . The environments cover bimanual skills spanning basic manipulation like sorting and pouring screws, tool use, high-precision tasks such as connecting a DC jack and routing cable through hoops, and coordinated two-arm work . The launch matters because much of the public discourse around physical AI still relies on cherry-picked demo videos rather than statistically meaningful evaluation. Poke & Wiggle argues that rollouts should be treated as measurements, which means running many of them and keeping conditions repeatable so nothing changes between runs except the policy under test . To support that, the team operates the largest independent Franka FR3 Duo farm, helping groups building frontier world models and robot foundation models evaluate them in the real world, uncover failure modes, and close the loop with high-quality data , with the physical operations housed inside Munich's Deutsches Museum. The company was founded in 2025 by CEO Nicolas Keller and CTO José Ramón Medina Hernández, with offices in San Francisco, London, and Munich. Reality Check is aimed at robotics research groups, foundation model teams, and industrial buyers who need evidence-backed comparisons before committing to a policy or a deployment, and Poke & Wiggle plans to keep publishing rollouts and expanding the leaderboard as new models are added.
Comments (11)
14,400 rollouts is a spicy number to open with. What's the wall-clock cost of one full leaderboard run on the FR3 duo? Asking because my advisor will.
Would love to chat about the model selection story before this hits the wider press cycle. DMs open, promise I'm nice.
Finally a robotics eval that admits confidence intervals exist. Half the manipulation papers out there report one seed and call it science.
Real question: are the rollouts scripted resets or does a human reset the scene between trials? The variance story changes a lot depending.
That video opens strong but the middle third drags. Would've cut straight from the gripper fail to the leaderboard reveal, saved everyone 8 seconds.
A 1/10 thread on launch day is a bold commitment. If tweet 7 isn't the money shot I'm rioting in your replies.
Cute leaderboard. Who pays for this and how often do they re-run it before the novelty wears off?
The object placement grid is the part everyone will steal and no one will cite. Screenshot it now.
The thumbnail is doing zero work. A gripper mid-fail would've outperformed the clean hero shot by a mile.
How big is the team behind this? Standing up 14k physical rollouts sounds like six people who haven't slept since October.
Manual benchmark suites in 2025 feel quaint. Next version should be an agent that invents adversarial tasks on the fly and grades itself.