Back to feed
News Story
APriority83
DeepTech深科技
1 sources

Tsinghua and Berkeley Propose Continuous World Model, Robots No Longer Guess Frame by Frame

Researchers from Tsinghua University's Institute for AI Industry Research and UC Berkeley have introduced ODEWorld, a continuous world model based on physical-time flow. Instead of predicting discrete frames, it learns how the world evolves continuously, boosting average success rates on real robot tasks from 55% to 80%. This could enhance robots' ability to understand and act in dynamic environments.

SynthePulse Insight · AI deep readingMembers

Continuous World Model ODEWorld: Teaching Robots How the World Flows

Version 1 · 1 source

Tsinghua and Berkeley teams propose ODEWorld, using physical time flow instead of discrete frame prediction to boost real-robot task success from 55% to 80%, but continuous time is not the end.

  • ODEWorld is a joint effort by Tsinghua AIR and Berkeley BAIR, centered on Physical Time Flow (PT-Flow), learning how the world changes continuously rather than predicting frame by frame.
  • Across 4 real-robot tasks, adding ODEWorld's continuous intermediate goals raised average success from 55% to 80%.
  • In 64-frame long-horizon prediction, ODEWorld has a latency of 0.072 seconds, far below V-JEPA 2's 0.619 seconds and LDP's 3.953 seconds.
Open section navigationFrom Discrete to Continuous: ODEWorld's Paradigm Shift

From Discrete to Continuous: ODEWorld's Paradigm Shift

Current mainstream world models often rely on discrete next-step prediction, slicing the continuous world into static frames, which leads to error accumulation and hallucinated outcomes in long-horizon tasks. ODEWorld introduces Physical Time Flow (PT-Flow), learning how the world continuously evolves from the current state to a future state, rather than simply predicting the next frame.

ODEWorld uses pre-trained DINOv2 to extract features, decouples static background from dynamic information via dynamic representation disentanglement, and employs JVP to map image changes into a compressed space, learning the direction and speed of change. During prediction, an ODE solver accumulates changes to generate states at any time point, even reversing to infer the past.

This continuous modeling not only fills in intermediate frames but changes how world models describe change: from learning 'what comes after A' to learning 'how A continuously transitions to B'.

Free for now

Read the full analysis

3 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

This article's information primarily comes from DeepTech's interview with co-first author Liu Dongxiu and the paper's data, constituting second-hand reporting. All performance numbers and success rates are from that report and have not been independently verified; they should be treated as source claims rather than confirmed facts.

Sources for this version

  1. 清华联合伯克利提出连续世界模型，机器人不用再逐帧猜环境DeepTech深科技

Primary report

DeepTech深科技

Primary source