Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
GPT-Policy is a research framework for in-context robot learning, described in the paper "In-Context Robot Learning with VLM Agents". It equips a general-purpose vision-language model agent (the authors use a frozen commercial model they call GPT-6 Astra) with a context compiler that organizes task information and an execution harness for geometric grounding, action chunking and closed-loop decisions based on execution feedback. Model parameters are not updated.
The framework accepts five kinds of context: human video, robot demonstrations (optionally with action references), a target image, self-interaction history including failures, and live human interaction such as pointing. Embodiment-specific adapters translate the VLM's tool requests into robot commands.
The project page reports evaluation on ten real-world tasks and ten RoboDojo simulation tasks. Across four real-world demonstration tasks, mean success rose from 10% without demonstrations to 80% with human videos or robot demonstrations including action references. In simulation, one-shot GPT-Policy reached 72% mean success versus 26% for the direct VLM and 48% for a π₀.₅ + VLM hybrid. The authors note that reliable physical execution remains a challenge, and that the simulation tasks informed development and are not a held-out evaluation. Execution is slow, at several minutes per task.
It is a research project rather than a product, and the page lists no pricing. A YouTube summary says code is released, but no repository URL was provided in the material. Running it depends on access to a commercial VLM and a robot setup.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Framework for in-context robot learning in which a vision-language model lets a robot learn a task from a human video, a goal image or feedback on failed attempts, without retraining. Real-world tests went from failure to 2 of 3 successes on tasks like picking up a towel and unscrewing a bottle cap. Code is released.
Closest neighbours by meaning, not just by category.
Model · Skild AI
Skild AI's robotic foundation model that performs new tasks from a single video demonstration, with no fine-tuning.
Model
27B vision-language robot policy trained only in simulation, run zero-shot on a real bimanual robot.
Model · Generalist AI
Robot foundation model from Generalist AI that learns new tasks in-context from a short demonstration.
Library
Embodied harness from NUS Show Lab that lets vision-language models control robots via discrete semantic action units.
Model
Research model that turns a pretrained video diffusion model into a latent action learner for human-to-robot motion transfer.
Model · Google DeepMind
Google DeepMind family of robot models: a VLA, ER 2 for reasoning and planning, and On-Device 2 for offline use.
