AI

H2R-Bench: Pretty Robot Videos Still Fail the Transfer

Shar Hendrix 4 min read

Egocentric human video is cheap. Robot tape is not. H2R-Bench, from Shanghai Jiao Tong University and Shanghai AI Lab (arXiv August 13), asks whether today’s video world models can turn one into the other.

Each case is a 5-second human clip, a target embodiment (parallel-jaw gripper or dexterous hand), and a prompt. The model must emit a robot video of the same task. Pose copy is not required. Keeping the goal, the actions, the functional contact, and the requested body is.

H2R-Bench pipeline from egocentric human videos to scored robot generations
240 transfer cases, six manipulation families, five scoring axes. Source: H2R-Bench project page.

240 cases, five scores

The bench uses 120 EgoDex clips, 20 in each of six families: rigid rearrangement, mechanism actuation, insertion, deformable configuration, bulk-material transfer, and surface change. Two embodiments make 240 cases.

Judges (Gemini 3.5 Flash, Qwen3.7-Plus, GPT-5.4) score 25 sampled frames on:

  • M1 Goal (weight 0.15)
  • M2 Action events (0.15)
  • M3 Functional contact (0.30)
  • M4 Embodiment (0.30)
  • M5 Video quality (0.10)

H2RCore is 100 times the weighted sum. Contact and body type carry 60% of the grade on purpose.

Three human raters line up with the automated transfer scores (within-scene Spearman ρ = 0.883, aggregate Pearson r = 0.930).

Qualitative human-to-robot video transfers on H2R-Bench
Qualitative transfers. A polished clip can still keep human hands or the wrong end-effector. Source: H2R-Bench project page.
Scatter of video quality versus H2RCore across models
Video quality sits in a 0.73–0.81 band while H2RCore runs 30.0 to 84.6. Source: H2R-Bench project page.

Who actually transfers

Video-conditioned models sit on top. Seedance 2.0 leads at 77.3 H2RCore on the gripper and 84.6 on the hand. Wan2.7 is next (76.5 / 83.1), then Kling-V3 (74.5 / 81.7).

Frame-conditioned systems drop hard on embodiment. Veo 3.1 still recognizes the task (gripper goal 0.725) but embodiment is 0.100 / 0.227. HunyuanVideo 1.5-I2V wins raw quality (0.806 / 0.808) and finishes last on H2RCore (30.0 / 30.7). Across 22 model–embodiment pairs, quality and H2RCore barely rank together (Spearman ρ = 0.14).

The dexterous hand is the easier target for 9 of 11 models, by +3.3 H2RCore on average. Giving Seedance nine frames instead of the full clip costs 27.6 gripper points and 41.4 hand points.

A Human’s Take

This is the right rude question. I do not want another “robot-looking” clip that still has human knuckles in it. If you are going to train a policy on generated tape, embodiment and contact are the whole product. H2RCore putting those two at 60% is the first scoring rule in this corner that matches how a robot actually fails.

Sources