Jia-Yu Chen

Human tactile gloves costing under $11 can now generate robot-ready tactile data, slashing acquisition costs by over 95% and boosting efficiency 3.5x for dexterous manipulation.

UCA-Flow achieves a remarkable 9.3% increase in action success rates while being up to 45.6 times faster than previous approaches.

EcoFrame achieves a remarkable 13.5x inference speedup while maintaining accuracy, revolutionizing how we approach long-video understanding in VLMs.

TRA achieves up to 2.05x faster video generation by intelligently mapping query entropy to token-specific attention budgets, revolutionizing efficiency in VDiTs.

EchoCache achieves a remarkable 2.46x speedup in audio-driven video generation without sacrificing quality or coherence.

Achieving 99.1% accuracy in object manipulation without the need for depth or additional 3D modules could revolutionize how robots interact with complex environments.

Regularizing the SAIL objective with reverse KL divergence not only resolves convergence issues but also enhances performance in LLM alignment tasks.

EcoVideo achieves up to 2.9x faster video generation in resource-constrained environments by intelligently orchestrating cloud-edge dynamics.

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

LargeMonitor achieves robust, zero-shot drift detection and dynamic adaptation in online continual learning, outperforming traditional methods that are blind to the structural causes of data shifts.

Forget short-sighted compression: Future Forcing anticipates future query needs in autoregressive video generation, boosting long-horizon consistency by up to 1.49 on VBench-Long without any training.

Real-world robots forget how to fold towels after learning to pick-and-place, but this work shows experience replay can help, if you do it right.

Extracting temporal geometry from generative models can boost reinforcement learning performance by over 2x without changing the optimal policy.

By explicitly modeling spatial interaction and manipulation intent, AIM achieves a 94% success rate on RoboTwin 2.0, suggesting that robot control can be drastically improved by explicitly reasoning about *where* to interact, not just *how* scenes evolve.