H2R-Bench (Video-Conditioned) - Parallel-Jaw Gripper: leaderboard

H2R-Bench (Rong et al., 2026) evaluates video world models that turn an egocentric human manipulation video into a robot manipulation video for a target embodiment, over 120 source videos in six manipulation families. Here the target is a parallel-jaw gripper. The H2RCore score from 0 to 100 weights goal-state completion, action-event completion, contact transfer and embodiment correctness, rated by three multimodal LLM judges, with video quality; this board covers models conditioned on the full source video, higher is better.

Metric: H2RCore (0-100) = 100 (0.15 goal-state completion + 0.15 action-event completion + 0.30 functional contact transfer + 0.30 embodiment correctness + 0.10 video quality); the first four are 0-4 evidence ratings by Gemini 3.5 Flash, Qwen3.7-Plus and GPT-5.4 on 25 sampled frames, averaged, and video quality combines MUSIQ, CLIP aesthetic, temporal-stability and AMT-S interpolation scores; 120 egocentric human manipulation videos in six families, each transformed into a robot manipulation video; the model receives the full source video; target embodiment: a parallel-jaw gripper. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=h2r-bench-video-conditioned-parallel-jaw-gripper · How It Works · Data refreshed daily, snapshot 2026-11-13.