Evidence-engineering guides · automated editorial preparation, not expert review
Robotics evidence and evaluation
What did a robot accomplish, under which conditions, with how much assistance—and what can another team actually verify? These guides develop evidence infrastructure for answering those questions. Humanoid shape alone is not proof of useful capability.
Forty topic pages cover foundations, evaluation, provenance, assistance, governance and worked examples. This is an evidence-engineering collection, not a safety assessment, certification service or hardware benchmark. The ROS and LeRobot walkthroughs are specifications, not operational adapters.
Start with a question
Foundations · 8 pages
- Robot embodiment is not a capability claim
- Write a robot task contract before collecting results
- Separate perception evidence from task completion
- What a pick-and-place result actually demonstrates
- Mobility results need localization context
- Separate plans, control and authority
- Keep human assistance visible in robot evaluations
- Robot evidence needs explicit frames and units
Evaluation · 10 pages
- Count all planned robot evaluation attempts
- Keep simulation and hardware evidence separate
- Replaying robot logs is not reproducing an experiment
- Compare controller revisions without moving the goalposts
- Choose a robotics benchmark for the decision it informs
- Report uncertainty in robot measurements
- Classify robot failures without inventing a cause
- Keep training exposure out of robot evaluation claims
- Document how a robot outcome was judged
- Turn a robot demonstration into an evidence-gap register
Provenance and data · 8 pages
- Trace robot datasets from collection to evaluation
- Align robot sensor time without hiding assumptions
- Bind robot measurements to calibration context
- Identify the model that actually ran
- Record the robot configuration behind each trial
- Separate access to robot episodes from permission to reuse them
- Retain enough robot evidence without keeping everything
- What a robotics artifact digest can and cannot prove
Human assistance · 6 pages
- Start assistive robotics with a bounded human need
- Evaluate access to the whole assistance workflow
- Measure supervised robotics as supervised robotics
- Treat home robot recordings as more than task data
- Record recovery without erasing the original failure
- Distinguish requesting a handoff from completing one
Governance · 4 pages
Specifications and examples · 4 pages
Relationship to the wider knowledge system
Use mathematics for formal foundations and governed workflows for the broader evidence boundary. Robotics applies those ideas to task observations; it does not turn a consistency check into physical approval.
Sources are linked with exact section locators. Workflows are Maha proposals; examples are synthetic unless explicitly stated otherwise. Search demand has not been measured for this collection.