Driving · 2026-10-14
RoboHarness: A Simple Harness Outperforms VLA and World Actions Models
Bince Qu, Wei Chen, Bo Zhang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
RoboHarness presents a robotic harness that gives LLM agents a visual-geometric control panel to directly invoke embodied tasks. Without model fine-tuning, the open-source model Qwen3.8-Flash-Next with RoboHarness is reported to outperform VLA and world action model baselines on BEHAVIOR Challenge 2025 tasks. The authors say the implementation will be released as open source.
Why it matters
The authors argue that the key limitation for embodied use of LLMs is the action interface to the physical world, not the LLM’s planning/spatial reasoning. RoboHarness aims to connect an unchanged LLM to robot control via tracked visual keypoints and geometric constraints, and reports improved performance versus VLA and WAM baselines on BEHAVIOR Challenge 2025.
Method
- Provide a robot harness that tracks marked visual keypoints, projects them with depth back-projection into a unified robot camera frame, and continuously returns coordinates to the LLM.
- Use a minimal, complete keypoint-constraint toolset: the LLM specifies geometric constraints for on-hand (modifiable) and off-hand (fixed) keypoints; the harness samples candidates, checks IK feasibility, plans trajectories, and executes.
- Add adjustment tools to directly command low-level robot degrees of freedom for refinement/robustness when constraints fail.
Limitation
To mitigate failures on BEHAVIOR long-horizon mobile manipulation tasks, the authors hand-craft a task-specific prompt for every task that specifies the execution steps and the behavioral boundaries the model must respect.
Abstract (from arXiv)
The rise of large language models (LLMs) has opened a path toward generalist embodied AI, raising the question of how LLMs can act in the physical world. One straightforward approach is training the model to generate actions: collect robot-action data and video demonstration, refine LLM architecture with an action head, and train a vision-language-action (VLA) model at scale. However, we argue that coding agents have the potential to realize generalist embodied manipulation. All they need is a simple yet effective robot harness. We present RoboHarness: a robotic harness that gives LLM agents a visual-geometric control panel so that they can directly understand and invoke embodied tasks. With RoboHarness, Qwen3.8-Flash-Next outperforms state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks. The implementation will be released as open source.
Related papers
- UniWALM: Unified World Action Language Model for Embodied Robot Control
- On Agentic Manipulation Policies
- EmbodiedEvo: Self-improving Coding Agents for Physical Intelligence
- LAW: Unifying Language, Action, and World Modeling for Agentic Physical Intelligence
- AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-09-29 · Indexed by Pub-AI: 2026-10-14 UTC
RoboHarness: A Simple Harness Outperforms VLA and World Action Models
Preprint posted: 2026-09-29 · Original record ↗
- arxiv
- 2610.14018
Explicit source versions (1)
- v2 · 2026-10-16
- Explicit source version detected · v2
Source date: 2026-10-16 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
- Preprint available
Source date: 2026-09-29 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
The rise of large language models (LLMs) has opened a path toward generalist embodied AI, raising the question of how LLMs can act in the physical world. One straightforward approach is training the model to generate actions: collect robot-action data and video demonstration, refine LLM architecture with an action head, and train a vision-language-action (VLA) model at scale. However, we argue that coding agents have the potential to realize generalist embodied manipulation. All they need is a simple yet effective robot harness. We present RoboHarness: a robotic harness that gives LLM agents a visual-geometric control panel so that they can directly understand and invoke embodied tasks. With RoboHarness, Qwen3.8-Flash-Next outperforms state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks. The implementation is open source.
