WorldBagel

“We introduce WorldBagel, a unified VLAW framework that extends BAGEL's two-tower GEN/UND architecture to jointly support multimodal understanding, action prediction, and future frame generation within a unified model.”

“WorldBagel achieves a 98.0% average success rate on the LIBERO benchmark, outperforming strong baselines like OpenVLA-OFT (97.1%) and π0.5 (86.8%).”

“FAST [35] tokenization relies on a BPE training procedure... making the tokenization data-dependent and potentially unstable across domains or control ranges. Their ablation in Table 4a shows FFAD achieving 98.0% success compared to FAST's 96.9%”

“π0.5 [21] ✓ ✓ 91.2 87.5 94.3 74.1 86.8”

“Co-author contributing to the unified VLAW formulation and theoretical proofs in Appendix A, ensuring that "errors in predicted Fourier features translate into bounded errors in reconstructed actions."”