AI Paper Atlas

Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Open-World Multi-Task Control Through Goal-Aware Representation Learning and Adaptive Horizon Prediction

Shaofei Cai, Zihao Wang, Xiaojian Ma, Anji Liu, Yitao Liang

CVPR 2023 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Imitation and skill learning · 10 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
use the full goal-sensitive backbone and adaptive horizon method instead of MT-BCMinecraft multi-task control · Plains, Flat, and ForestAvg. Success Rate and Avg. Precision: Plains: 32_{\pm05} -> \boldsymbol{55}_{\pm09} (success rate), 43_{\pm06} -> \boldsymbol{70}_{\pm09} (precision); Flat: 3Improved
add goal-sensitive backbone features (ablation)Ours and MT-BC · Plains, Flat, and Forest multi-task benchmarksSuccess Rate and PrecisionImproved
add adaptive horizon prediction (ablation)Ours · Plains multi-taskAvg. Success Rate: 35_{\pm08} -> \boldsymbol{55}_{\pm09}Improved
add a GRU module while retaining horizon prediction (ablation)Ours · Plains multi-taskAvg. Success Rate: 55_{\pm09} -> \boldsymbol{65}_{\pm07}Improved
add horizon prediction alongside a GRU module (ablation)Ours without horizon prediction + RNN · Plains multi-taskAvg. Success Rate: 39_{\pm08} -> \boldsymbol{65}_{\pm07}Improved
add horizon loss (ablation)Ours · Plains multi-taskAvg. Success Rate: 47_{\pm06} -> \boldsymbol{55}_{\pm09}Improved
include extra observation input (ablation)Ours · Plains multi-taskAvg. Success Rate: 50_{\pm07} -> \boldsymbol{55}_{\pm09}Improved
include language condition (ablation)Ours · Plains multi-taskAvg. Success Rate: 25_{\pm03} -> \boldsymbol{55}_{\pm09}Improved
subtract a small positive constant from predicted horizon at evaluation (ablation)Horizon-conditioned policy · Flat multi-taskPerformanceImproved
subtract a larger value from predicted horizon (ablation)Horizon-conditioned policy · Flat multi-taskPerformanceHurt
train on Plains and evaluate without fine-tuning on FlatOurs · Flat novel-biome evaluationAvg. Success Rate: 63 -> 58Mixed

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