From Human Videos to Robots: Early Zero-Shot Reward Model Explorations
JasonMa2020 · x · 2026-08-13
The author revisits early research on zero-shot transferring human-video-trained reward models to robots, noting that while the idea seems ancient by today's standards, it laid the groundwork for modern embodied AI.
- VIP Model: Cites the 2022 paper VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training. The research introduced a self-supervised pre-trained visual representation capable of generating dense and smooth reward functions for unseen robotic tasks using an implicit time contrastive objective.
- Technical Evolution: Mentions Jesse's Robometer work, emphasizing that using human videos to provide zero-shot capabilities for robot reward models has a long history, with recent approaches like Dyna-2 finally demonstrating scaling transfer on robot actions.
Related event: Human Video Data Enhances Robot Generalization(2 posts)→
More from Embodied
- Hardware Hack: Reviving Robot Bob with Dev Kits and Vision Sensors — chrismatthieu · 2026-08-13
- Unitree IPO Valuation Sparks Debate: Prediction Market Prices It at $35B — MarvinTBaumann · 2026-08-13
- Sci-Fi Made Real: Top VR Players to Pilot Physical Humanoid Robots in Combat — cixliv · 2026-08-13
- Gemini Robotics Tested: Humanoid Robot Shows Human-like Motor Generalization — keerthanpg · 2026-08-13
- Gripper Deformation Skews Robot Evals: A Hidden Hardware Trap — DominiqueCAPaul · 2026-08-13
- Hands-on with Unitree R1: A Cheap, Mini G1 Dev Platform — carlosdponx · 2026-08-13