AffectiveVLA: Safe and Intent-Aware VLA Control for Human-Humanoid Social Interaction
AffectiveVLA: Safe and Intent-Aware VLA Control for Human-Humanoid Social Interaction
This paper introduces AffectiveVLA, a Vision-Language-Action pipeline designed for safe social interaction between humanoids and humans. The system uses onboard cameras to infer interaction intent from RGB data, converting it into language prompts for the VLA policy. It features low-confidence suppression to prevent unintended motion execution and includes compliance profiles that adjust arm impedance during contact, alongside a voice-triggered safe-stop mechanism. The authors collected approximately 1,300 multimodal episodes via VR teleoperation to fine-tune four models: GR00T, RLDX-1, Being-H0.5, and UnifoLM-VLA-0. Evaluation on the Unitree G1 identified GR00T as the best standalone model, with intent recognition integration raising its macro-F1 score from 0.75 to 0.90 and achieving a p95 inference latency of 6.27 ms for the intent model.
The study presents AffectiveVLA, an intent-gated Vision-Language-Action pipeline evaluated on a Unitree G1 humanoid robot using VR-collected multimodal data and four fine-tuned VLA models.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
interaction episodes
≈1,300
Reported trials
View original evidence
ntent from two onboard cameras and converts it into a language prompt for the VLA policy. Low-confidence predictions suppress motion execution, reducing unintended activations. During physical contact, interaction-specific compliance profiles adjust arm impedance, while an independent voice-triggered mechanism initiates a latched safe-stop procedure. Using VR teleoperation, we collected approximately 1,300 multimodal
Open source S3inference latency (p95)
6.27
View original evidence
interaction episodes and fine-tuned four VLA models: GR00T, RLDX-1, Being-H0.5, and UnifoLM-VLA-0. On-robot evaluation identifies GR00T as the best-performing standalone model. Integrating intent recognition increases its macro-F1 score from 0.75 to 0.90, while the intent model achieves a p95 inference latency of 6.27 ms. These results demonstrate that explicit intent gating improves the reliability of VLA-based soc
Open source S4training data size
≈1,300
Reported trials
View original evidence
ntent from two onboard cameras and converts it into a language prompt for the VLA policy. Low-confidence predictions suppress motion execution, reducing unintended activations. During physical contact, interaction-specific compliance profiles adjust arm impedance, while an independent voice-triggered mechanism initiates a latched safe-stop procedure. Using VR teleoperation, we collected approximately 1,300 multimodal
Open source S3
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2611.07712v1 Announce Type: new Abstract: Social touch interactions, such as handshakes, fist bumps, and hugs, require humanoid robots to infer human intent, generate contact-rich motions, and maintain safe physical contact. We present AffectiveVLA, an intent-gated Vision-Language-Action (VLA) pipeline deployed on a Unitree G1 humanoid with dexterous hands. A lightweight RGB-based model recognizes interaction i
Open source S2
ntent from two onboard cameras and converts it into a language prompt for the VLA policy. Low-confidence predictions suppress motion execution, reducing unintended activations. During physical contact, interaction-specific compliance profiles adjust arm impedance, while an independent voice-triggered mechanism initiates a latched safe-stop procedure. Using VR teleoperation, we collected approximately 1,300 multimodal
Open source S3
interaction episodes and fine-tuned four VLA models: GR00T, RLDX-1, Being-H0.5, and UnifoLM-VLA-0. On-robot evaluation identifies GR00T as the best-performing standalone model. Integrating intent recognition increases its macro-F1 score from 0.75 to 0.90, while the intent model achieves a p95 inference latency of 6.27 ms. These results demonstrate that explicit intent gating improves the reliability of VLA-based soc
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org