HKU team launches The Imitator Game: a 4-level benchmark for intent-level robot imitation
YiMaTweets · x · 2026-09-04
A team from the University of Hong Kong, TranscEngram, Fudan and Zhejiang universities released The Imitator Game, a hierarchical Human→Robot video imitation benchmark asking whether robots can imitate human intent rather than just replay actions.
- Building on prior work (Generalist AI, Skild AI) showing human-video prompting enables one-shot robot task learning, the benchmark targets the next question: how far can robots go beyond trajectory replay?
- Four levels: L0 scene-identical execution → L1 spatial adaptation → L2 visual/physical generalization → L3 intent-level functional transfer, progressively widening the gap between demonstration and the robot's own scene.
- Released resources: 20K+ paired real/sim episodes, 200+ task variants and simulation environments, evaluations of 9 models / 15 variants, plus a human evaluation leaderboard (Imitator Arena).
Paper, code, dataset and leaderboard are publicly available.
More from Embodied
- Unitree brings robot demo to IFA show floor — melnykowycz · 2026-09-04
- Tesla's Cybercab is real — now everything hinges on scaling, says stock watcher — JOBhakdi · 2026-09-04
- Tesla Opens Cybercab Robotaxi Rides to the Public Tomorrow at 5pm CT — aelluswamy · 2026-09-04
- E-ink bag charm TICKEY raises ~$380K on Kickstarter in two weeks — 创业邦 · 2026-09-04
- Sam Altman confirms OpenAI will build its own humanoid robots, aiming for one per person — 创业邦 · 2026-09-04
- Figure Humanoid Robot Adds Stair Climbing Skills for Industrial Deployment — Anen-o-me · 2026-09-04