dWorldEval: Scalable Robotic Policy Evaluation via Discrete Diffusion World Model
| Authors | Yaxuan Li et al. |
| Year | 2026 |
| HF Upvotes | 1 |
| arXiv | 2604.22152 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Evaluating robotics policies across thousands of environments and thousands of tasks is infeasible with existing approaches. This motivates the need for a new methodology for scalable robotics policy evaluation. In this paper, we propose dWorldEval, which uses a discrete diffusion world model as a scalable evaluation proxy for robotics policies. Specifically, dWorldEval maps all modalities - including vision, language, and robotic actions - into a unified token space, modeling them via a single transformer-based denoising network. In this paper, we propose dWorldEval, using a discrete diffusion world model as a scalable evaluation proxy for robotics policy. Specifically, it maps all modalities, including vision, language, and robotics action into a unified token space, then denoises them with a single transformer network. Building on this architecture, we employ a sparse keyframe memory to maintain spatiotemporal consistency. We also introduce a progress token that indicates the degree of task completion. At inference, the model jointly predicts future observations and progress token, allowing automatically determine success when the progress reaches 1. Extensive experiments demonstrate that dWorldEval significantly outperforms previous approaches, i.e., WorldEval, Ctrl-World, and WorldGym, on LIBERO, RoboTwin, and multiple real-robot tasks. It paves the way for a new architectural paradigm in building world simulators for robotics evaluation at scale.
Engineering Breakdown
Plain English
This paper addresses a critical bottleneck in robotics: evaluating policies across thousands of environments and tasks is computationally infeasible with current methods. The authors propose dWorldEval, a discrete diffusion world model that acts as a fast, scalable proxy for policy evaluation without running expensive simulations. The key innovation is mapping vision, language, and robotic actions into a unified token space, then using a single transformer-based denoising network to model all modalities jointly. This enables efficient evaluation of robotics policies at scale by predicting policy outcomes synthetically rather than requiring real or high-fidelity simulated rollouts.
Core Technical Contribution
The core novelty is using a discrete diffusion model as a learned world model specifically designed for scalable policy evaluation in robotics. Instead of training separate models for vision, language, and action, dWorldEval unifies all modalities into a single discrete token vocabulary and denoises them through a transformer architecture. This is fundamentally different from prior world models (like video prediction networks or action-conditioned simulators) because it's optimized for evaluation speed and accuracy trade-offs rather than pixel-perfect reconstruction. The discrete tokenization approach enables efficient batching and inference compared to continuous diffusion models, making it practical for evaluating policies across thousands of tasks.
How It Works
dWorldEval takes a policy and environment description (including vision, language context, and action history) as input and tokenizes all modalities into a unified discrete token space using learned codebooks or vocabulary mappings. A transformer-based denoising network is trained to predict the next state given the current noisy state representation, following the discrete diffusion framework where tokens are progressively corrupted during training and the model learns to reverse this corruption. During evaluation, the system conditions on the policy's action and iteratively denoises to predict future states, effectively simulating policy rollouts without actually executing them. The discrete formulation allows efficient batching of multiple trajectories and tasks simultaneously, and the unified token space means a single model handles all modalities rather than separate pathways. The output is a predicted trajectory or outcome that indicates whether the policy would succeed on that task, enabling rapid filtering and ranking of policies without expensive simulation.
Production Impact
For robotics teams, this approach could reduce policy evaluation time from hours or days to minutes, enabling faster iteration cycles during development and hyperparameter tuning. Instead of running full simulations on a simulator farm or waiting for real robot time, engineers could use dWorldEval as a cheap pre-filter to identify promising policies before expensive validation. The unified token-space architecture simplifies the model pipeline — you train one model instead of maintaining separate vision encoders, language models, and action decoders, reducing engineering complexity and potential failure modes. The main trade-off is accuracy: the world model is an approximation, so policies that pass evaluation may still fail in reality, requiring validation on actual deployments. Integration requires collecting training data of robot trajectories, policy actions, and outcomes, which means you need a substantial database of successful and failed rollouts before deployment.
Limitations and When Not to Use This
The approach assumes the discrete diffusion model can accurately capture the complex dynamics of robotic environments, but significant distribution shift between training data and novel environments could cause evaluation failures to go undetected. The method requires a large dataset of diverse trajectories to train the world model effectively, making it less practical for teams with limited robotics data or novel task categories not represented in training. The paper doesn't clearly address how well the tokenization generalizes to continuous control subtleties or how errors propagate across long-horizon predictions — multi-step denoising could accumulate errors that invalidate the entire trajectory. There's also an open question about how well this scales to real-world high-dimensional tasks with visual perception and complex language instructions, especially when the training distribution doesn't cover rare failure modes the policy might encounter.
Research Context
This work builds on decades of world model research in RL (like Dreamer and Latent World Models) but applies the discrete diffusion framework that has become mainstream in vision and language (following diffusion models' success in image generation and recent work like LLaMA tokenization). It's motivated by the scalability crisis in robotics benchmarking — existing evaluation approaches like BenchBot and RoboNet require either real robots or high-fidelity simulators, making broad policy comparison infeasible. The discrete diffusion choice is particularly interesting given recent success in multimodal learning with unified token vocabularies (similar to approaches in video generation and embodied AI). This opens a research direction toward learned evaluation functions that could generalize across task families and robot morphologies, potentially enabling a 'policy benchmark leaderboard' where policies are evaluated on thousands of tasks using a single learned proxy model.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::