ai-trend-notifier
← wiki

$ cat wiki/people/jim-fan.md

Jim Fan

person

Overview

NVIDIA Senior Research Scientist. A leading researcher in Embodied AI / Foundation Agent. Leads Project GR00T (humanoid robot foundation model). High signal frequency at the intersection of agents + robotics — 1.2x weight.

Affiliations

  • NVIDIA (Senior Research Scientist, AI Agents)

Notable Recent Statements

  • 2026-06-17: ENPIRE — agentic robot policy self-improvement in the real world — Jim Fan's NVIDIA GEAR Lab (with CMU, UC Berkeley) released ENPIRE: a fleet of 8 real robots autonomously runs its own research loop — no human researchers in the loop. Results: 99% pass@8 on contact-rich tasks; physical scaling law discovered (8 parallel robots → superlinear policy improvement). Fan: "AutoResearch in the physical world for the first time." Builds directly on EgoScale (June 7): EgoScale → acquire skills from human video; ENPIRE → autonomously improve those skills. → ENPIRE: Agentic Robot Policy Self-Improvement in the Real World (source) (arXiv)

  • 2026-06-07: EgoScale — humanoid dexterous manipulation from egocentric human video — Announced that his NVIDIA team trained a humanoid with 22-DoF dexterous hands (Sharpa Wave tactile + Unitree H2 Plus, 75 DoF total) to assemble model cars, operate syringes, sort cards, and fold shirts — all learned from 20,000+ hours of egocentric human video with no robot teleoperation. Key finding: log-linear scaling law (R² = 0.998) between human video volume and real-robot task success rate — the strongest signal yet that LLM-era scaling extends to dexterous manipulation. Also demonstrated live VR teleoperation (PICO headset → Unitree G1). Open-sourced everything. (source)

  • 2026-04: CaP-X open-sourced — "Vibe agents alive in the physical world", robot arms + humanoids, perception/actuation APIs, auto-synthesize skill libraries (source)

  • Foundation Agent roadmap: a single model learns to act across diverse virtual/physical worlds — Project GR00T as the cornerstone

  • No-gradient architecture stance: "LLM acts as 'prefrontal cortex' that orchestrates lower-level control APIs via code generation" (citing Voyager)

Key Themes

  • Embodied AI / robotics
  • Foundation agents (cross-environment)
  • LLM as high-level controller, with low-level control kept separate
  • Open-source first

Why Tracked

  • Represents the agentic robotics direction of NVIDIA (TBD)
  • High frequency of agents matches + real-world application signals
  • Intellectual parallel to Karpathy's AI-native software vision (abstracting humans-environments)

Related

Referenced by