Contracts before execution
A policy declares cameras, state format, control rate, and action conventions. The resolver records a direct, adapted, or refused model-benchmark pairing before simulation.
physical intelligence · benchmarked
The benchmarking platform for physical intelligence.
pera helps teams building physical AI evaluate VLA and robot-policy models under reproducible benchmark protocols, with pinned tasks, seeds, scoring rules, episode videos, action traces, and verification-ready evidence.
pinned protocols
per-episode evidence
verification commands
how pera works
From physical-AI model submission to verifiable result.
Every pera evaluation follows a visible, reproducible workflow.
why pera is different
The score stays attached to the evidence.
pera is built for researchers, model builders, and evaluation agents expanding physical AI who need to inspect why a result happened, not just cite the aggregate number.
Failures stay inspectable
Result pages expose task-level scores, failure labels, episode videos, review coverage, and whether a label was manually reviewed or inferred from a task pattern.
Verification has clear limits
Env-verified rows attest environment execution, declared protocol, scoring, transcripts, and evidence integrity. Signed model-identity attestations are a later tier.
Conflict-of-interest disclosure: PulseVLA is verapulse's own model. Its rows use the same protocols, evidence surface, and disclosure rules as the other models.
benchmarks and compatibility
Supported protocols, model rows, and evidence status.
Open a cell to inspect the full row. Protocol cards show the simulator, episode budget, step cap, task list, and number of evaluated models.
- models · - protocols · - published result rows
Select any score to inspect the full result page.
reproduction reports
What ordinary leaderboards miss.
The reports document exact replications, quarantined runs, convention bugs, and cross-model effects that are easy to hide inside a single score.
report no. 1 · exact replication
10 / 10 tasks
OpenVLA-OFT matched the authors' harness exactly
Episode-count parity also exposed a quaternion branch bug that dropped one task from 98% to 58%.
Read report → report no. 2 · rejected run 84.4%A plausible score was rejected
Episode evidence revealed corrupted observations in 52 of 98 failures. The clean rerun scored 95.2%.
Read report → report no. 4 · contract failures 3 bugsSilent convention mismatches stopped before publication
Rotation, proprioception frame, and image orientation all produced clean execution and wrong behavior.
Read report → report no. 5 · cross-model finding 4 / 4Every eligible model showed the same resolution effect
The strongest task-level gain was 26 additional successes out of 50, with no direction reversals.
Read report → report no. 3 · honest baseline 27%Octo's unflattering result reproduced too
A verification system earns trust by reproducing low scores and failure structure, not only flattering numbers.
Read report →same rules, your model
Put your physical-intelligence model under the same evaluation protocol.
Connect a compatible policy, run a supported benchmark, inspect every episode, and publish a result another team can verify.
Need help wiring an adapter or interpreting a result? Join the pera Discord for community support.