AIAny
Icon for item

HumanCLAW: Can Vision-Language Models Act Through a Body?

Evaluates whether vision-language models can make actionable decisions for a physical body by decoupling decision-making from low-level motor execution. Introduces HumanCLAW-Bench with 1,218 long-horizon egocentric episodes across 41 indoor scenes and diagnoses a lack of embodied self-awareness in current VLMs.

Introduction

Why this matters Vision-language models (VLMs) are increasingly tasked with embodied roles, but real-world action couples high-level decisions with noisy motor control — failures blur whether fault lies in the model's choice or execution. HumanCLAW sidesteps this confound by harnessing off-the-shelf VLMs to emit atomic skill commands that are deterministically translated into short, physically realistic full-body motion chunks. This isolates the model's action intelligence (what it chooses next) from execution errors (balance, motor noise), enabling clearer evaluation of embodied decision-making.

Key Findings
  • A focused evaluation protocol: HumanCLAW decouples decision from execution by using a harnessed pipeline where VLM outputs map to sub-second full-body motions that still experience gravity and collisions, so choices have real physical consequences while execution-side stochasticity is minimized.
  • New large benchmark: HumanCLAW-Bench contains 1,218 long-horizon find-navigate-interact episodes spanning 41 indoor scenes, emphasizing egocentric perception and multi-step planning over extended trajectories.
  • Quantitative diagnosis: Nine state-of-the-art VLMs are evaluated and none solve the benchmark; the best model achieves only a 16.8% success rate, showing that perception/recognition alone is not the bottleneck.
  • Core limitation identified: Contemporary VLMs lack embodied self-awareness — they struggle to track their own body, recognize whether they reached goals, and detect collisions or obstacles in the loop of action selection.
Who it's for and trade-offs

Great fit if you need a principled way to evaluate high-level action decisions from multimodal models without conflating results with low-level motor failures. The benchmark is useful for researchers building embodied VLM agents, designers of action grounding methods, and teams developing evaluation suites for long-horizon loco-manipulation tasks. Look elsewhere if your primary interest is improving raw motor controllers or hardware-level balance and actuation; HumanCLAW intentionally factors out execution noise and therefore does not evaluate controller robustness or actuation improvements.

Where it fits

HumanCLAW sits between perception benchmarks (which test recognition) and full end-to-end robotics evaluations (which conflate perception, planning, and control). It provides a middle ground: physically plausible action outcomes with minimized execution confounds, making it a diagnostic tool to advance embodied decision-making in multimodal models.

Information

  • Websitearxiv.org
  • AuthorsSiyao Li, Jiawei Gu, Shuai Liu, Kairui Hu, Zekun Li, Linjie Li, Chengcheng Tang, Po-Chen Wu, Ivan Shugurov, Lingni Ma
  • Published date2026/07/29

More Items

Evaluates multimodal context learning across grounding, new information application, and knowledge acquisition using a 3,443-instance benchmark spanning science, finance, long documents, spatial reasoning, and web VQA; finds current multimodal models perform poorly (best score 0.2847) and analyzes failure modes.

Directly maps visual observations and language instructions to continuous robot actions, replacing LLM-centric V→L→A pipelines. Uses separate visual and language encoders with lightweight bidirectional interaction and a compact decoder to cut inference cost and VRAM, achieving ~31 ms latency and <1 GB VRAM on an RTX 4090; suited for real-time robotic manipulation under tight compute budgets.

Real-time streaming multimodal foundation model that uses a codec-native tokenizer (Mage-ViT) to encode motion- and residual-rich regions from video I/P frames, reducing visual token usage by over 75% and enabling up to ~3.5× wall-clock inference speedup after training on ~560M images and 100M video frames.