imitation (HumanCompatibleAI)

Action Chunking with Transformers (ACT) Redefines the Economics of Bimanual Manipulation

Tony Z. Zhao, Chelsea Finn, and colleagues introduce Action Chunking with Transformers (ACT), an algorithm that enables low-cost, imprecise hardware to achieve 80-90% success rates in fine-grained bimanual tasks using just 10 minutes of teleoperation data. By addressing compounding errors through generative action sequence modeling, ACT shifts the industry focus from expensive sensor suites to efficient imitation learning pipelines, fundamentally altering the hardware economics for dual-arm robo

Diffusion Policy Redefines Visuomotor Imitation Learning with 46.9% Performance Leap

Diffusion Policy reframes robot control as conditional denoising of action sequences, achieving a 46.9% average performance improvement across 12 manipulation benchmarks. By leveraging stochastic Langevin dynamics and time-series diffusion transformers, it effectively resolves multimodal action distributions and high-dimensional space challenges. The public release of its code and data provides a stable, open-source backbone for embodied AI startups targeting complex physical deployment.