Embodied harness from NUS Show Lab that lets vision-language models control robots via discrete semantic action units.
Updated EU status checked Last shown
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.
Show-Harness is a research framework from Show Lab at the National University of Singapore that lets vision-language models (VLMs) "play" robots through a compact semantic interface. The VLM picks one named, parameter-free action unit per step (for example MV_LEFT), and an embodiment-specific interpreter deterministically converts it into local robot motion. The same vocabulary is reported to drive a 7-DoF Franka, a 6-DoF AgileX and a simulator without changing the model.
The project describes two routes: zero-shot control with closed-source frontier VLMs (e.g. Gemini-3.1 Pro), and low-cost fine-tuning of small open-source backbones such as Qwen3.5-2B and InternVL3.5 with a few GPU-hours. According to the project page, both outperform VLA baselines (π0.5, GR00T), VLA-centric agents and code-as-policy approaches. Ablations indicate that the naming and convention of action units matter heavily, and that frontier VLMs fail mainly on centimetre-level spatial grounding rather than planning.
It also includes GUMI (GUI Manipulation Interface), which maps action units to labeled keys so a human or GUI agent can collect demonstrations without specialized teleoperation hardware. Results are demonstrated on real robots and in simulation; using it for physical control requires a robot. The provided material did not include a code repository link, license or paper link.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Harness that lets any vision-language model (Gemini, GPT, Qwen, GLM) control robots by choosing from named actions, converted by an action interpreter. Success reached up to 100%; requires a real robot.
Closest neighbours by meaning, not just by category.
Model
27B vision-language robot policy trained only in simulation, run zero-shot on a real bimanual robot.
Model · Perceptron
Perceptron's vision-language model for physical AI, covering perception, physical reasoning and robot action prediction.
Model
Action-conditioned robotics world model fine-tuned from Cosmos-Predict2.5-2B using 2D skeleton conditioning across embodiments.
Model · Unitree
Unitree's 6B-parameter vision-language-action foundation model for humanoid robots, covering 64 tabletop and whole-body tasks.
Research
Research framework letting a frozen VLM agent learn robot tasks in context from videos, images and feedback, no fine-tuning.
Model · Xiaomi
Open-weights Vision-Language-Action robot foundation model from Xiaomi, pre-trained on 100K+ hours of UMI trajectories.