TLDR: Hybrid Training (HyT) is a new framework for Vision-Language-Action (VLA) models that allows robots to learn from complex ‘Chain-of-Thought’ reasoning during training, internalizing knowledge for improved performance. Crucially, HyT enables the VLA to then execute actions directly and quickly during real-time operation, avoiding the inference slowdown typically associated with generating intermediate thoughts. This approach delivers high performance and fast inference, making VLAs more practical for real-world robotic tasks.
In the rapidly evolving field of robotics, Vision-Language-Action (VLA) models are paving the way for more generalist robots. These advanced models take language instructions and camera images as input, then output low-level robotic actions, enabling robots to perform complex tasks. However, a common challenge with these powerful models, especially those using ‘Chain-of-Thought’ (CoT) reasoning, has been a trade-off between performance and speed.
Traditional CoT strategies, where a VLA generates intermediate ‘thoughts’ before taking an action, have significantly boosted performance. This is similar to how humans might consciously deliberate before acting on a complex problem. While these ’embodied CoT’ (ECoT) methods improve a robot’s ability to understand and execute tasks, the generation of these intermediate thoughts adds to the model’s processing time, slowing down the robot’s actions. In real-world scenarios, particularly in robotic manipulation, delays can severely impact usability.
A new research paper introduces an innovative approach called Hybrid Training (HyT) that aims to resolve this dilemma. The core idea behind HyT is to allow VLAs to learn from these valuable ‘thoughts’ during training, internalizing the knowledge and performance benefits, without necessarily needing to generate them during real-time operation. This means the robot can still act quickly, much like a human developing ‘skilled intuition’ – where complex decisions become effortless over time due to learned patterns.
How Hybrid Training Works
HyT enables a single VLA model to learn multiple ways of generating outputs, depending on a ‘modality variable’. During training, the model is exposed to a mix of data, learning to:
- Act Directly: Like a standard VLA, predicting actions immediately.
- Think First: Similar to ECoT, generating intermediate thoughts before actions.
- Follow Instructions: Acting as a low-level policy, following provided thoughts or instructions (e.g., from a human or another system).
By learning from this diverse set of objectives, the model internalizes a deeper understanding of tasks and environments. Crucially, at inference time, the model can be instructed to operate in an ‘act’ mode, directly predicting actions without generating intermediate thoughts. This allows HyT-trained VLAs to maintain the same fast inference speed as standard VLAs, while still benefiting from the knowledge acquired through CoT training.
The flexibility of HyT also means the model can still be used in ‘think’ mode for interpretability (to understand the robot’s intentions) or ‘follow’ mode for fine-grained instruction following, offering a versatile tool for robotic control.
Also Read:
- Enhancing Multimodal Reasoning with Advanced Vision-Language Process Reward Models
- Evaluating Vision-Language-Action Models for Real-World Industrial Use
Real-World Impact and Performance
The researchers rigorously tested HyT across various simulated benchmarks, including ClevrSkills and LIBERO, and in real-world robotic experiments. The results consistently showed that HyT-trained models not only outperform standard VLAs but also generally surpass models trained with ECoT and hierarchical VLA methods across different data scales. This performance boost was particularly evident in more complex tasks and in scenarios requiring generalization to new, out-of-distribution settings.
In real-world tests using a UFactory xArm 6, HyT demonstrated superior performance compared to OpenVLA, especially in tasks involving novel objects or placements. The HyT-trained robot showed greater precision in reaching picking and placing positions, avoiding common errors like reaching for the wrong object.
This research highlights that the true value of Chain-of-Thought techniques for VLAs lies in the enhanced understanding and representation learning they provide during training. By internalizing this reasoning, HyT allows robots to achieve higher performance with faster execution, making them more practical and efficient for real-world applications. For more details, you can read the full research paper here.
