Benchmarks

Task NameDataset NameSOTA ResultTrendResults

Robot Co-design

Hopper search phase

1.51D2C Score

Offline Reinforcement Learning

hopper medium-replay v2

366Return

Noisy Action Robust MDP

Hopper high randomness MuJoCo Noisy Action Robust MDP

331.69Worst-case Return

Noisy Action Robust MDP

Hopper MuJoCo Noisy Action Robust MDP medium randomness

482.2Worst-case Return

Noisy Action Robust MDP

Hopper MuJoCo Noisy Action Robust MDP (low randomness)

477.9Worst-case Return

Reinforcement Learning

Hopper

998.99Average Decisions

Continuous Control

Hopper

2.51Action Repetition

Reinforcement Learning

Hopper

400,000Episode Count

Meta-Reinforcement Learning

Hopper params

0.51FLOPs (k)

Policy Improvement

Hopper H

92Success Rate

Continuous Control

Hopper (meta-test)

485Total Reward

Continuous Control

Hopper (meta-train)

2,825Total Reward

Continuous Control

Hopper v5

270.9Terminal Performance

Reinforcement Learning

Hopper v5

270.9Terminal Performance

Reinforcement Learning

Hopper standard (test)

24.6Episode Length

Interpretability Evaluation

Hopper

3.1Interpretability Score

Continuous Control

Hopper v4 (test)

206.36Mean Return

Locomotion

Hopper

—Primary metric

Hopper

—Primary metric