The authors designed FutureWorlds, built on a multimodal discrete autoregressive model.
Robotic world models learn by comparing diverse predicted futures
FutureWorlds uses diverse beam search and candidate-specific memory to generate informative alternative futures, then trains via group-relative advantage to improve prediction quality.
Chinese Tech
Hao Wu · Shengju Qian · Weiyan Wang · Fan Xu · Fan Zhang · Yuanpeng He · +2 more
HKUST (GZ) · CUHK · Tencent · USTC · PKU
Research Digest··3 min read
The authors propose FutureWorlds, a framework for learning robotic world models by constructing, maintaining, and comparing multiple alternative futures.
Why this paper
From Tencent and 5 others
In one line
FutureWorlds reduces LPIPS by 14.78%, 20.84%, and 9.12% on three datasets for 32-frame predictions.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§