Tsinghua AI

Extracting interaction cues from a frozen video model enables robots to achieve up to 90.6% success in manipulation tasks without costly rollout processes.

HALO-WA boosts robotic manipulation success rates from 26.4% to 87.1% by effectively adapting to real-world errors in just over an hour of training.

State-of-the-art planners falter in long-tail scenarios, revealing critical gaps in autonomous driving safety and effectiveness.

Bridge-WA achieves superior task performance by predicting where and how the world will change, enabling robots to focus on relevant scene dynamics rather than irrelevant visual details.

WorldSample achieves a 28% boost in policy success rates while slashing training steps by nearly 60% through innovative real-synthetic data integration.

A single BDDL specification can drastically enhance the efficiency and effectiveness of embodied task planners, achieving a 25.9% performance improvement over existing baselines.

Current interactive world models fall short, with none passing the rigorous tests of WorldRoamBench designed to assess long-horizon stability across action, vision, physics, and memory.

PhysEditWorld reveals that explicit control over physical parameters can transform how game world models interact with their environments, leading to more realistic and manipulable simulations.

Teacher-forcing consistency models can accelerate autoregressive video generation by ten times, revolutionizing the training landscape for streaming applications.

ASSCG cuts inference latency by 60% while boosting performance scores in autonomous driving systems, redefining how LLMs can be efficiently integrated into fast-slow planning architectures.

Counterfactual controllability in video generation could be the key to creating self-evolving world models that understand and adapt to their actions.

WVM outperforms existing models by accurately assessing task progressions and improving robotic manipulation from both expert and suboptimal data.

H-RePlan achieves a remarkable increase in task completion rates by intelligently distinguishing between local and global recovery strategies in multi-device environments.

MemoryWAM achieves superior performance in robotic manipulation tasks by efficiently leveraging both short-term and long-term memory without sacrificing computational efficiency.

ImageWAM shows that image editing can outperform video generation in robot action prediction, cutting costs and improving efficiency.

By harnessing implicit supervision from environment dynamics, EnvRL boosts RL success rates by over 4% on long-horizon tasks, revealing a new frontier in agentic learning.

DRIVE-CHOREO achieves unprecedented multi-view consistency and BEV mAP by choreographing latent tokens across diverse modalities in autonomous driving.

Jointly optimizing the world model and action model is essential for mastering long-horizon tasks, revealing a critical gap in traditional WA training methods.

PearlVLA achieves state-of-the-art performance in action generation by refining plans in latent space, enabling low-latency execution without sacrificing deliberation quality.

Advanced planners still exhibit critical safety failures, with FluidTest uncovering new threats in over 65% of evaluated trajectories.

HABC achieves up to 92% success in complex robotic tasks by intelligently balancing viability and efficiency in sparse outcome scenarios.

Simulated evaluations can mislead policy rankings, but our findings reveal how to better align simulations with real-world performance.

Envision4D transforms future scene extrapolation in autonomous driving, achieving unprecedented accuracy and stability in dynamic environments.

Hierarchical latent actions in HiMem-WAM boost long-horizon manipulation robustness, outperforming existing models in real-world scenarios.

Treating raw visual images as action representations revolutionizes embodied control, outperforming traditional methods in accuracy and generalization.

Raw context outperforms compact memory designs, revealing that memory structure is crucial for effective video generation in action-conditioned models.

Reliable civil court judgments can now be simulated with a framework that adapts to the complexities of legal claims and remedies.

CP4D achieves photorealistic 4D scene generation by seamlessly integrating static environments with dynamic objects, outperforming existing methods in visual fidelity and physical consistency.

Generating realistic 3D environments from satellite imagery in under 10 minutes could revolutionize how we visualize and interact with our planet.

Text world models can transform LLM-based agents from reactive responders into proactive planners, enhancing their performance in complex interactive tasks.

AnchorWorld's innovative use of 3D human motion and exogenous viewpoints enables a new level of interaction fidelity in egocentric simulations, setting a new benchmark in the field.

Achieving 93.7 on the NAVSIM v1 benchmark, CLEAR redefines the efficiency of multi-modal planning in autonomous driving without the need for complex iterative processes.

WorldFly's innovative approach to integrating world models enables UAVs to navigate complex urban landscapes with unprecedented robustness.

Bridging the perception-reasoning gap in visual planning, MGSD boosts model performance by over 19% while relying solely on visual inference during deployment.

Real-time infinite video generation is now feasible, achieving over 1.3 million frames in a single 24-hour rollout without sacrificing quality.

GIM-World achieves superior long-horizon visual consistency by integrating geometry-aware implicit memory, outperforming traditional memory systems.

Forget fixed agent slots and quadratic attention: Gamma-World uses simplex embeddings and sparse hubs to generate interactive multi-agent environments with better fidelity and control, even generalizing from 2 to 4 players without retraining.

Embodied agents get a 30%+ performance boost by using external tools for perception and reasoning, but still struggle to pick the right tool for the job.

Unlock scalable autonomous driving simulation with AnyScene, a framework that generates controllable, high-fidelity driving scenes from arbitrary BEV layouts and camera configurations.

LLMs can dynamically replan sub-goals during embodied instruction following, leading to state-of-the-art performance on ALFRED.

Unleash parallel robot learning with MuJoCoUni, enabling high-throughput batched physics evaluation without sacrificing upstream MuJoCo semantics.

Current mobile GUI agents struggle with complex, long-horizon tasks in realistic simulated environments, achieving only a 17.82% success rate on SimuWoB.

Forget tedious expert demonstrations: this RL approach learns autonomous parking by smartly incorporating human corrections and failed attempts, achieving impressive real-world performance.

One-Forcing achieves state-of-the-art one-step video generation while slashing training costs to a third of previous methods.

Stop letting bad actions ruin your VLA rollouts: Pre-VLA uses runtime verification to boost success rates by 7% while slashing execution steps.

Quadrupedal robots can now perform dynamic loco-manipulation in the real world, matching human teleoperation, using only onboard ego-centric vision and a low-frequency (5Hz) open-vocabulary detector.

By pretraining a VLA model with goal-conditioned RL, PRTS learns to reason about goal reachability, leading to substantial gains in long-horizon robotic tasks and zero-shot generalization.

Achieve real-time robotic action with 79-91% success while generating high-fidelity 4D reconstructions, all within a single unified world model.

Autonomous vehicles can now plan trajectories 10x faster without sacrificing performance, thanks to a novel architecture that learns complex driving behaviors in latent space during training.

Achieve superhuman dexterity: ALAS unlocks robust long-horizon task completion by decoupling environment understanding from motor control, enabling generalization across diverse human-scene interaction scenarios.