AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

PlayVirtual: Augmenting Cycle-Consistent Virtual Trajectories for Reinforcement Learning

Tao Yu, Cuiling Lan, Wenjun Zeng, Mingxiao Feng, Zhizheng Zhang, Zhibo Chen

NeurIPS 2021 · Reinforcement learning › Visual and representation RL · approach: Latent-space world models · 8 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
augment SPR with cycle-consistent virtual trajectoriesSPR on Atari-100k · Atari-100kmedian HNS (%): 37.1 -> 47.2Improved
augment SPR† with cycle-consistent virtual trajectoriesSPR† on DMControl-100k · DMControl-100kmedian score: 728.0 -> 797.0Improved
augment SPR† with cycle-consistent virtual trajectoriesSPR† on DMControl-500k · DMControl-500kmedian score: 920.0 -> 935.0Improved
add future-state prediction (ablation)SPR on Atari-100k · Atari-100kmedian HNS (%): 33.4 -> 37.1Improved
add BDM without virtual-trajectory augmentation (ablation)SPR on Atari-100k · Atari-100kmedian HNS (%): 37.1 -> 38.4No clear effect
add cycle-consistent virtual trajectories (ablation)Baseline+BDM on Atari-100k · Atari-100kmedian HNS (%): 38.4 -> 47.2Improved
allow cycle loss to update the dynamics model as well as the encoder (ablation)PlayVirtual-ND on Atari-100k · Atari-100kmedian HNS (%): 44.0 -> 47.2Improved
set prediction steps to K=12 instead of K=9 (ablation)PlayVirtual on Atari-100k · Atari-100kmedian HNS (%): 47.2 -> 43.1Hurt
calculate cycle distance in projection space rather than latent space (ablation)PlayVirtual on Atari-100k · Atari-100kmedian HNS (%): 44.8 -> 47.2Improved
remove cycle-loss weight warmup (ablation)PlayVirtual on DMControl-100k · DMControl-100kmedian score: 749.5 -> 797.0Hurt
apply cycle constraint at every trajectory step instead of only the end step (ablation)PlayVirtual on DMControl-100k · DMControl-100kmedian score: 781.0 -> 797.0Hurt
increase virtual trajectories from M=20 to M=30 (ablation)PlayVirtual on DMControl-100k · DMControl-100kmedian score: 806.0 -> 792Hurt

Later papers that compared with this method

4 later accepted papers report PlayVirtual as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy