Robbyant Releases LingBot-VLA 2.0: An Open-Source 6B Vision-Language-Action (VLA) Model for Cross-Embodiment Robot Manipulation
Robbyant releases LingBot-VLA 2.0, a 6B open-source Vision-Language-Action model based on Qwen3-VL-4B, targeting cross-embodiment generalization. The model utilizes a unified 55-dimensional canonical action space and a sparse Mixture-of-Experts (MoE) architecture to scale efficiently across diverse robot types. Training leverages 60,000 hours of curated data (robot trajectories and egocentric videos) filtered via explicit smoothness metrics and VLM-based annotation. Predictive dynamics are enhan
Analysis
TL;DR
- Robbyant releases LingBot-VLA 2.0, a 6B open-source Vision-Language-Action model based on Qwen3-VL-4B, targeting cross-embodiment generalization.
- The model utilizes a unified 55-dimensional canonical action space and a sparse Mixture-of-Experts (MoE) architecture to scale efficiently across diverse robot types.
- Training leverages 60,000 hours of curated data (robot trajectories and egocentric videos) filtered via explicit smoothness metrics and VLM-based annotation.
- Predictive dynamics are enhanced through dual-query distillation from depth and video teacher models, allowing the agent to anticipate future states.
- Benchmarks on GM-100 and long-horizon mobile manipulation tasks show significant improvements in success rates, particularly in out-of-distribution scenarios compared to previous versions.
Why It Matters
This release addresses the critical industry challenge of deploying VLA models outside controlled laboratory environments by emphasizing robust generalization across different robot embodiments. By providing an open-source, Apache-2.0 licensed model with a standardized action representation, it lowers the barrier for developers to implement generalist robot policies without needing to train from scratch for each specific hardware configuration. The focus on predictive dynamics and high-quality data filtering offers a practical blueprint for improving real-world robotic reliability and performance.
Technical Details
- Architecture: Built on the Qwen3-VL-4B-Instruct backbone with a 6B parameter checkpoint. The action expert employs a sparse Mixture-of-Experts (MoE) design with SwiGLU MLPs, using a sigmoid-based routing strategy inspired by DeepSeek-V3 to balance load without auxiliary losses.
- Unified Action Space: Employs a fixed 55-dimensional canonical vector to represent states and actions across 20 different robot configurations, including arms, hands, grippers, waists, heads, and mobile bases, with padding for missing components.
- Data Pipeline: Curated 60,000 hours of pre-training data (50k robot hours, 10k egocentric hours). Filtering involves computing third-order jerk and Z-scores for velocity/acceleration, removing static or noisy episodes, and using Qwen3.6-27B for automated subtask segmentation and atomic action labeling.
- Predictive Dynamics: Uses two learnable queries (current and future state) supervised by distillation from LingBot-Depth (geometric cues) and DINO-Video (semantic priors based on DINOv3 with block-wise causal temporal attention).
- Performance: Achieves ~130ms inference latency on an RTX 4090D. Outperforms baselines like GR00T N1.7 and pi0.5 on the GM-100 bimanual benchmark and shows strong in-domain and out-of-distribution results on long-horizon tasks like refrigerator sorting and stove cleaning.
Industry Insight
- Standardization of Action Spaces: The adoption of a unified canonical vector for diverse embodiments suggests a growing industry trend toward hardware-agnostic robot policies, reducing integration complexity for multi-robot fleets.
- Data Quality Over Quantity: The rigorous filtering pipeline using physical metrics (jerk, smoothness) and VLM-based annotation highlights the importance of high-quality, curated datasets in achieving robust generalization, moving beyond simple data volume accumulation.
- Efficiency via MoE: The successful application of auxiliary-loss-free MoE routing in robotics indicates that sparse activation techniques can effectively scale model capacity while maintaining low-latency inference required for real-time robotic control.
Disclaimer: The above content is generated by AI and is for reference only.