Hi WorldEngine team,
While reading the paper (Section 3.3, Eq. 6–8) and cross-referencing it with Nexus (Zhou et al., ICCV 2025, cited as [16] in the paper), I noticed strong architectural overlap between the Behaviour World Model (BWM) and Nexus:
The two papers share 3 authors (Tianyu Li, Naisheng Ye, Hongyang Li).
WorldEngine Eq. 6–7 defines the noise schedule as k = [k_a, τ] ∈ (0,1]^{A×T}, where each k_{a,τ} is an independent per-token noise level — this matches Nexus's core "decoupled diffusion" contribution (independent noise states for fine-grained tokens), rather than standard full-sequence denoising or next-token prediction.
BWM's "goal conditioning" mechanism (Sec. 2.3) also appears consistent with Nexus's goal-state guidance.
BWM's "optimization guidance" mechanism (Sec. 2.3) seems to correspond to a separate cited work ([17], "Optimization-guided diffusion for interactive scene generation") rather than to Nexus itself.
Given this, I'd like to ask:
Is BWM's diffusion transformer core literally the Nexus architecture (or a direct fine-tune/extension starting from the publicly released nuplan.ckpt), or is it an independently-trained model that happens to share the same architectural design?
Is BWM effectively "Nexus's decoupled-diffusion core + optimization guidance from [17]" integrated together, or are there additional architectural differences not described in the paper?
Is there any plan to release the BWM generator itself (model code + checkpoint used to produce the data/sim_engine/scenarios/augmented/* scenarios)? Currently the repo only ships a consumer for pre-generated augmented scenarios (NavSimOpenSceneE2EFineTuneSynthetic), not the generator.
Were the publicly released augmented/navtrain_50pct_collision (and ep_1pct/offroad) scenarios generated using the exact same BWM checkpoint/pipeline that produced the paper's reported 5,340-scene / 31,508-frame post-training dataset (Sec. 4.1), or from a different/smaller internal run?
Thanks for the great work and for open-sourcing such a comprehensive codebase!
Hi WorldEngine team,
While reading the paper (Section 3.3, Eq. 6–8) and cross-referencing it with Nexus (Zhou et al., ICCV 2025, cited as [16] in the paper), I noticed strong architectural overlap between the Behaviour World Model (BWM) and Nexus:
The two papers share 3 authors (Tianyu Li, Naisheng Ye, Hongyang Li).
WorldEngine Eq. 6–7 defines the noise schedule as k = [k_a, τ] ∈ (0,1]^{A×T}, where each k_{a,τ} is an independent per-token noise level — this matches Nexus's core "decoupled diffusion" contribution (independent noise states for fine-grained tokens), rather than standard full-sequence denoising or next-token prediction.
BWM's "goal conditioning" mechanism (Sec. 2.3) also appears consistent with Nexus's goal-state guidance.
BWM's "optimization guidance" mechanism (Sec. 2.3) seems to correspond to a separate cited work ([17], "Optimization-guided diffusion for interactive scene generation") rather than to Nexus itself.
Given this, I'd like to ask:
Is BWM's diffusion transformer core literally the Nexus architecture (or a direct fine-tune/extension starting from the publicly released nuplan.ckpt), or is it an independently-trained model that happens to share the same architectural design?
Is BWM effectively "Nexus's decoupled-diffusion core + optimization guidance from [17]" integrated together, or are there additional architectural differences not described in the paper?
Is there any plan to release the BWM generator itself (model code + checkpoint used to produce the data/sim_engine/scenarios/augmented/* scenarios)? Currently the repo only ships a consumer for pre-generated augmented scenarios (NavSimOpenSceneE2EFineTuneSynthetic), not the generator.
Were the publicly released augmented/navtrain_50pct_collision (and ep_1pct/offroad) scenarios generated using the exact same BWM checkpoint/pipeline that produced the paper's reported 5,340-scene / 31,508-frame post-training dataset (Sec. 4.1), or from a different/smaller internal run?
Thanks for the great work and for open-sourcing such a comprehensive codebase!