The field of robotics is moving towards the development of more generalizable and robust Vision-Language-Action (VLA) models. Recent research has focused on improving the performance of VLA models in various tasks, such as visual navigation, manipulation, and tracking. One of the key directions is the integration of multimodal learning, where models are trained on multiple sources of data, including vision, language, and action. This approach has shown promising results in improving the generalization capabilities of VLA models. Another important area of research is the development of more efficient and scalable training methods, such as reinforcement learning and meta-learning. These methods have been shown to improve the performance of VLA models in complex tasks and environments. Noteworthy papers in this area include MM-Nav, which proposes a multi-view VLA model for robust visual navigation, and TrackVLA++, which enhances embodied visual tracking with spatial reasoning and temporal memory mechanisms. Overall, the field is moving towards the development of more advanced and generalizable VLA models that can be applied to real-world robotic applications.
Advances in Vision-Language-Action Models for Robotics
Sources
From Pixels to Factors: Learning Independently Controllable State Variables for Reinforcement Learning
MM-Nav: Multi-View VLA Model for Robust Visual Navigation via Multi-Expert Learning
Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer
Fine-Tuning on Noisy Instructions: Effects on Generalization and Performance
LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization
Seeing the Bigger Picture: 3D Latent Mapping for Mobile Manipulation Policy Learning
Bridge Thinking and Acting: Unleashing Physical Potential of VLM with Generalizable Action Expert
SITCOM: Scaling Inference-Time COMpute for VLAs
NoTVLA: Narrowing of Dense Action Trajectories for Generalizable Robot Manipulation
Zenbo Patrol: A Social Assistive Robot Based on Multimodal Deep Learning for Real-time Illegal Parking Recognition and Notification
ContextVLA: Vision-Language-Action Model with Amortized Multi-Frame Context
Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time
HyperVLA: Efficient Inference in Vision-Language-Action Models via Hypernetworks
StaMo: Unsupervised Learning of Generalizable Robot Motion from Compact State Representation
Focused Skill Discovery: Learning to Control Specific State Variables while Minimizing Side Effects
Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment
RLinf-VLA: A Unified and Efficient Framework for VLA+RL Training
Bring the Apple, Not the Sofa: Impact of Irrelevant Context in Embodied AI Commands on VLA Models
Vision-Language-Action Models for Robotics: A Review Towards Real-World Applications
Generative World Modelling for Humanoids: 1X World Model Challenge Technical Report
TrackVLA++: Unleashing Reasoning and Memory Capabilities in VLA Models for Embodied Visual Tracking
TIGeR: Tool-Integrated Geometric Reasoning in Vision-Language Models for Robotics