HOST (Human-to-robot One-Shot Skill AcquisiTion) is a July 2026 arXiv cs.RO preprint from Beijing Institute of Technology, X Square Robot, and Tsinghua University. The authors treat new-skill teaching as inference-time video conditioning instead of teleoperation plus fine-tuning, which they say is slow and overwrites prior skills.
That matters for builders of vision-language-action and world-action systems because it would add skills the way retrieval adds documents: freeze the weights, store a video. The newsroom question was what actually transfers, and on which hardware. Primary sources show a narrower result than a cross-robot one-shot product: human-to-robot transfer of held-out tasks on one dual-arm cell, after training on hundreds of thousands of same-robot trajectories, with no reported simulation suite.
What we checked
- Does HOST acquire a novel manipulation skill at inference from a single human video without updating policy weights? confirmed. After two-stage pretraining, HOST conditions a frozen dual-expert diffusion policy on one human video and reports 62 percent average success on a novel-task comparison subset, about 29 seconds per skill including recording. New skills stay in an external video store rather than in weights, so previously mastered tasks are retained. [1][3][2]
- Does HOST transfer one-shot skills across different robot platforms and report simulation baselines? refuted. Every reported robot experiment uses one bimanual ARX R5 platform. The authors state HOST remains untested on robots that differ substantially in embodiment. The paper, GitHub, and project site describe no Isaac Gym, MuJoCo, or other simulation evaluation. [1][2][3]
What we found
- HOST does not map a human video directly to actions. A Qwen3-VL-Embedding-8B aligner trained with temporal cycle-consistency and Smooth DTW places the human video and a robot trajectory on a shared task-progress manifold, then each prediction target is the robot future that matches upcoming video progress. A dual-expert autoregressive diffusion model, with a video expert initialized from Wan2.2-TI2V-5B, then localizes progress, predicts future robot observations, and emits an action chunk. That cascade is the claimed one-shot mechanism. [1][2]
- All reported physical experiments use two ARX R5 six-axis arms with parallel-jaw grippers, three RGB cameras, and a 20-dimensional Cartesian-plus-gripper action space. HOST is evaluated on 50 novel manipulation tasks, 20 trials each, judged by humans against task-specific criteria. The authors write that evaluations are on a single bimanual platform and that HOST remains untested on robots that differ substantially in embodiment. No simulation baseline appears in the paper, project site, or repository. [1][3]
- Inference-time one-shot sits on a large same-embodiment pretrain. Stage 1 uses 193,462 robot trajectories across 229 tasks as robot-robot pairs. Stage 2 adapts on 5,847 self-collected human videos paired with same-task robot trajectories. GitHub reports paper-scale training on 64 GPUs: 10,000 alignment steps, 500,000 Stage 1 steps, and 100,000 Stage 2 steps. Ablations show Stage 2 pairs alone are insufficient; novel-task success rises monotonically with Stage 1 data volume. [1][2]
- On the comparison subset the project site lists eight tasks at 50 to 80 percent success, averaging 62 percent. HOST exceeds the strongest one-shot visual-imitation baseline by 43 percentage points on that table (62 versus 19 for AWDA) and exceeds zero-shot language VLAs by 45 percentage points as stated in the paper. Wall-OSS with LoRA on 50 teleoperated demos per task reaches 56 percent. Acquisition is about 29 seconds including recording, versus about 4.0 hours for pi0.5 plus 50-demo fine-tuning, stated as 507 times faster. The strongest fine-tuned baseline retains 43 percent of prior-task performance; HOST keeps weights frozen. [1][3]
- The public checkpoint is not the human-to-robot system in the paper. Hugging Face Guangyan/HOST is a roughly 15.9 GB robot-robot self-grounded predictor; the model card says task-video conditioning comes from peer robot episodes, not human video. GitHub eval_openloop.sh replays recorded episodes and does not command a live robot. Paper-scale datasets are not shipped, and training configs still point at internal cluster paths. [4][2][5]
- X Square Robot's 3 August 2026 news restates 62 percent success and 507 times faster acquisition, but collapses the 29-second acquisition time into a single 29-second video and describes the method as predicting task outcomes rather than imitating motion. That is a company paraphrase of self-grounded future-observation prediction, not an independent measurement. No peer-reviewed venue or third-party robot study was found in this session. [6][1]
What it means
Treat HOST as evidence that a frozen, progress-aligned world-action model can add manipulation skills from a human video without LoRA, not as a cross-embodiment one-shot product. If you copy the architecture, budget Stage 1 same-robot paired data at the 100,000-trajectory scale, a smaller human-robot adaptation set, Wan2.2-class video compute, and an external video memory. Do not ship the public Hugging Face file as the paper's human-to-robot policy, and do not assume it will run on a different arm, gripper, or action space without new paired data. Keep the 62 percent figure tied to the eight-task comparison subset and human-judged trials on one ARX R5 cell. Video also omits contact force, which the authors flag as a limit for fine-grained manipulation.
Still open
- Has any lab reproduced the 62 percent novel-task result, or is a peer-reviewed venue attached to arXiv 2607.20033?
- Will the authors release Stage 2 human-to-robot weights and the 193,462 plus 5,847 episode corpora, or only the robot-robot checkpoint?
- How semantically far are the 50 novel tasks from the 229 training tasks, and what are per-task rates on the full set of 50?
- Can the same cascade transfer to a second robot morphology without collecting a new human-robot pair set?