Two Meta-questions about Benchmarks

• What is a benchmark testing? (i.e., are these the right metrics)
  • Nature is not single-objective
  • Solving a specific task vs designing general purpose systems

• Is this benchmark representative of real-world problems?
  • Are potential advances useful in the real-world?
  • If using simulations, is the result indicative?

Solving a specific task vs designing general purpose systems

• In Robotics there is tension between creating systems that just work, and advancing scientific understanding
• Do we care about being able to pick the same object over and over ?
  (e.g., industrial application)
• Or do we care about a system that can adapt to different tasks (potentially unknown at training time)?
• In System Identification (and kids), we do not know the task beforehand. Can we still learn something useful?

Benchmarks on Real Robots are Hard...

• Not everyone has the same setting (robot, sensors, etc)
• Running real-world experiments is time-consuming and expensive
• Can we just use simulation?

Formalizing the Hardware

Final Remarks

• What are we really benchmarking? (Perception vs learning vs controller vs hardware)
• Agree with Jan's point about "mine is better than yours"
• Reproducing and negative results MUST be worth (see Physics)
• Many of the current learning benchmarks (e.g., OpeanAI) are atrocious
  (clearly not well designed as meaningful benchmarks)
• Benchmarks should not be proprietary (e.g., MuJoCo)

Mimic Benchmark

• Let a Robot "play" in an environment for a long time (e.g., 3 Months) without any goal
• Now bring in humans and the robot has to reproduce any skill that the human demonstrate
• The humans win if the robot can not reproduce the shown skill
  (Generative Adversarial Human)
• How long does it take to the Human to win?