HOST Framework: Robots Learn Household Tasks 500x Faster from 29s Video
APPSO · wechat · 2026-08-03
Self-Variable Robotics, collaborating with BIT and Tsinghua, open-sourced HOST (Human-to-robot One-Shot Skill Acquisition). This framework allows dual-arm robots to execute new tasks simply by watching a short human demo video, without updating model parameters.
- Efficiency Leap: Compared to traditional data collection and fine-tuning (which takes 4 hours), HOST reduces skill acquisition time by roughly 507x, succeeding across 50 completely new tasks.
- Core Tech: Using a dual-expert MoT architecture, HOST avoids rigidly mimicking human motion. Instead, it aligns progress in a vector space, allowing the robot to "imagine" the completed task and reverse-engineer the required actions.
- No Catastrophic Forgetting: Because learning happens entirely during inference without altering base parameters, the robot retains old skills and can dynamically replan actions if items are moved.
Papers, code, and models are fully open-sourced, offering a promising path for consumer robots to learn directly from ordinary users at home.
More from Embodied
- Reproducing FACTR 2 Force Feedback: Impedance Control Enables Safer Robot Interaction — RemiCadene · 2026-08-03
- Elevator-LIO: Open-Source LiDAR-Inertial Odometry for Robust Cross-Floor Localization in Elevators — rsasaki0109 · 2026-08-03
- Honor Robot Phone Review: Mechanical Gimbal + AI Brain Turns Phones into Companions — 数字生命卡兹克 · 2026-08-03
- CS-JEPA: Decentralized Predictive Architecture for Swarm Robotics — ITMO · 2026-08-03
- ODEWorld: A Continuous-Time Latent World Model via ODE — Dongxiu Liu · 2026-08-03
- N_0-TWAM: First Large-Scale Tactile-Native World-Action Model for Robots — NeoteAIEmbodied · 2026-08-03