WorldDiT — Unified Diffusion for Robotics World Modeling and Control (Bagel Labs)

TL;DR

Bagel Labs announces WorldDiT, a single diffusion transformer backbone that jointly samples robot actions and predicts a future observation of the world in parallel — no separate VLM in the control loop. Framed as sitting on the reported Pareto frontier of the LIBERO manipulation benchmark and as the best publicly released method among those that do not need a VLM to generate actions. At inference the policy acts for a few steps, observes what changed, and replans; the world-modeling branch is training-only, so deployment stays action-only and light-weight. Announcement-only tweet — no arxiv, blog, or code linked; no per-task numbers or parameter count given in the thread.

Key claims

WorldDiT unifies action sampling and future-frame prediction inside one diffusion backbone; both are learned in parallel during training [tweet thread §1–3].
WorldDiT does not require a large pretrained VLM to produce actions, cutting billions of parameters out of the control loop [tweet thread §2].
On LIBERO, WorldDiT is the strongest publicly released method that does not use a VLM to generate actions, and its (size, performance) pair sits on the Pareto frontier of reported results [tweet thread §1].
At inference, the world-modeling branch is dropped: the policy acts for a few steps, observes, and replans, so deployment is action-only [tweet thread §4].

Method

WorldDiT is described in the thread as a single diffusion transformer that consumes an observation and denoises jointly over (a) a robot action chunk and (b) a predicted future observation, sharing gradients between the two heads through one backbone. The world-modeling signal is presented as the richer supervision that lets the model hit strong LIBERO results with far fewer parameters than VLA-with-VLM baselines. A comparison table on LIBERO is shown as an embedded image; specific rows, model sizes, and success rates are not transcribed in the thread text and no report is linked at filing time.

Results

Positioned as best among publicly released VLM-free methods on LIBERO and on the reported Pareto frontier [tweet thread §1, embedded LIBERO table image].
No per-task numbers, LIBERO-suite splits (Spatial / Object / Goal / Long), parameter counts, or baseline model versions were transcribed in the tweet text; the LIBERO comparison table is only shown as an image.
No real-robot results claimed.

Why it’s interesting

WorldDiT stakes out a specific point on the VLA Models recipe board that the wiki has been tracking: it drops the VLM backbone entirely, so it’s not a “VLM + action head” (Spirit-v1.5, Embodied-R1.5, LingBot-VLA) nor a “frozen VLA under an orchestrator” (Harness VLA, VoLo). Structurally it’s closest to the frozen action-free world-model + small action expert recipe of μ₀: A Scalable 3D Interaction-Trace World Model (µ₀) and the unified-video+action-diffusion recipe of Causal World Modeling for Robot Control (LingBot-VA) (LingBot-VA), but pushes further: instead of factoring WM and action, WorldDiT bets that a single diffusion transformer denoising both future observations and actions in parallel is the right unification — with the world-modeling branch dropped at inference so control latency is action-only. Complements 1XWM: From Video to Action — 1X's video-pretrained world model as a NEO robot policy (video WM is the policy, IDM extractor) as another VLM-free-VLA data point: two different labs converged on “no VLM in the loop” in the same week via very different factorizations. This is also the second Bagel Labs artifact filed after Paris 2.0: A Decentralized Diffusion Model for Video Generation / Paris: A Decentralized Trained Open-Weight Diffusion Model — a lab known for decentralized video-diffusion training now shipping a robotics-policy artifact from the same architectural lineage.