AIAny
Icon for item

Show-Harness: Just a VLM Agent Can Play Robots

Lets vision-language models control robots via a compact semantic action interface that maps intent to discrete action units; supports zero-shot use of closed-source VLMs, low-cost fine-tuning of open VLMs, and GUI-based demonstration collection.

Introduction

Most robotic control work either trains embodiment-specific policies or end-to-end models that fuse perception and low-level control. Show-Harness takes a different tack: it treats foundation vision-language models (VLMs) as high-level decision-makers and exposes a small, semantic action language that the VLM can reason over, while deterministic embodiment-specific interpreters ground those semantics into actual robot motions.

Key Findings
  • Semantic action abstraction: Exposes a compact set of discrete action units (directional moves, grasp/release, terminal) so VLMs reason at a symbolic level instead of low-level trajectories — this isolates high-level planning from embodiment details.
  • Deterministic grounding: Uses embodiment-specific interpreters to convert semantic units into local robot commands, keeping the VLM responsible for fine-grained decisions while ensuring repeatable, hardware-safe execution.
  • Practical unlocks: Demonstrates two deployment modes — zero-shot control of closed-source frontier VLMs and rapid adaptation of small open VLMs with only a few GPU-hours — lowering the barrier to real-robot use.
  • GUI demonstration pipeline: GUMI extends the same semantic action space to GUI-based demo collection, enabling humans and agents to "play" multiple embodiments without specialized teleoperation hardware.
  • Empirical behavior: Reported experiments show robust generalization across tasks, embodiments, and environments, with performance gains over representative agentic and VLA baselines.
Who it's for and trade-offs

Great fit if you want to leverage existing VLMs as planners for manipulation tasks, rapidly prototype multi-embodiment agents, or collect demonstrations without robot teleops. Look elsewhere if your primary need is millimeter-level control, extremely low-latency closed-loop control, or tasks that cannot be meaningfully abstracted into discrete semantic units — the approach trades end-to-end low-level control for modularity and generalization.

Where it sits

Conceptually positioned between end-to-end vision-to-action systems and fully modular tool-based harnesses: the harness provides a small, model-agnostic semantic interface that unlocks large pretrained VLM reasoning for embodied tasks while relying on deterministic grounding modules for embodiment-specific safety and precision.

Information

  • Websitearxiv.org
  • AuthorsYanzhe Chen, Zechen Bai, Zhijun Cao, Wenzheng Zeng, Kevin Qinghong Lin, Yiqi Lin, Guoqiang Liang, Kevin Yuchen Ma, Qiming Huang, Mike Zheng Shou
  • Published date2026/09/09

More Items

Defines the Discovery Certification Protocol (DCP) to audit claims of discovery by AI research agents, converting claims into executable recovery and feedback tests. Specifies multi-gate certification, Core control requirements, and a deterministic offline verifier; validated in two controlled audits.

Evaluates whether AI agents can autonomously use Sparse Autoencoders (SAEs) to discover interpretable features and causally steer model behavior. Introduces a 20-task benchmark that has agents design contrastive probes to search a 131K+ feature dictionary in Gemma-2-9B-IT and evaluates activation rank, selectivity, and causal steering against expert references.

Translates natural-language instructions into executable programs that maintain an explicit, persistent global world state and compiles state-augmented 3D oriented bounding boxes into pixel-aligned conditioning signals for pretrained video generators. The approach decouples state evolution from rendering, enabling programmable entity control, off-screen state, and long-horizon interactive scenarios.