TL;DR
This paper introduces "Torque-aware Vision-Language-Action" (TA-VLA) models, enabling robots to better understand and react to physical interactions by integrating joint torque data. They found that incorporating torque information into the decoder, using historical torque data as a single summarized token, and predicting future torques alongside actions significantly improves performance, especially in contact-rich tasks.
What Problem Does This Paper Address?
Robots performing manipulation tasks, especially those involving physical contact (like plugging in a charger or assembling parts), need to understand how they are interacting with the environment. Joint torque, a measure of the forces and torques experienced by a robot's joints, provides rich information about contact dynamics that is often imperceptible from visual or language inputs alone. However, existing Vision-Language-Action (VLA) models, which aim to bridge perception, language, and action, typically struggle to effectively utilize this crucial physical feedback. This limits their ability to perform complex, contact-rich tasks reliably and robustly. The authors aim to bridge this gap by systematically exploring how to best integrate torque signals into VLA architectures.
Methods
The authors systematically investigate the "design space" for incorporating torque signals into VLA models. This involves exploring three key dimensions:
- When: Whether to use immediate, historical, or future torque data.
- Where: Whether to integrate torque into the encoder (processing sensory inputs) or the decoder (generating actions).
- How: Whether to use single-frame torque, multiple historical torque frames, or a summarized historical torque token.
They propose and evaluate several strategies, including:
- Torque Adapters: Using MLPs to process torque signals before integrating them into the VLA architecture.
- Decoder Integration: Placing torque information within the decoder, arguing it aligns better with other proprioceptive signals and leverages the decoder's sensitivity to fine-grained variations.
- Historical Torque Summarization: Encoding the entire history of torque signals into a single token to preserve decoder input patterns while retaining information.
- Unified Action-Torque Diffusion Model: A novel approach that jointly predicts future actions and torques, encouraging the model to build a physically grounded understanding of interaction dynamics.
Datasets
The paper does not explicitly mention the creation of new datasets. Instead, they fine-tune existing VLA models (π0, RDT, ACT) on their collected dataset, which includes 10 diverse tasks (5 contact-rich and 5 regular). They also mention using publicly available pre-trained weights for their baseline models.
Key Insights and Ideas
- Decoder is King for Torque: Integrating torque signals into the decoder of a VLA model consistently outperforms integrating them into the encoder. This is because torque is a proprioceptive signal that aligns better with other joint-related information already processed by the decoder, and the decoder is more sensitive to these subtle variations.
- History Matters, but Summarized: Torque history is far more informative than just the current frame. However, feeding multiple historical torque frames can disrupt the decoder's learned input patterns. The most effective approach is to summarize the entire torque history into a single token, balancing informativeness with architectural stability.
- Predicting the Future (Torque) Pays Off: By training the model to predict future torques alongside future actions (inspired by autonomous driving planning), the model develops a more physically grounded internal representation of interaction dynamics. This auxiliary task significantly boosts performance.
Key Results
The proposed TA-VLA model, particularly the "π0+obs+obj" variant (combining torque observations and the joint action-torque prediction objective), achieves significant performance gains:
- Contact-Rich Tasks: The combined approach consistently outperforms strong VLA baselines (ACT, RDT, π0) on contact-rich manipulation tasks, with success rates often reaching near-perfect levels (e.g., 18/20 for Button Pushing, 17/20 for Charger Plugging).
- Generalization: The torque integration strategies also improve performance on regular tasks, demonstrating their broad utility.
- Cross-Model and Cross-Embodiment: The benefits of torque integration generalize across different VLA architectures (RDT) and robotic embodiments (a different robotic arm), highlighting the robustness of the approach.
- Qualitative Improvements: Visualizations show the robot successfully retrying failed attempts due to misalignments or slippages by leveraging torque feedback, leading to task completion.
Limitations and Future Work
- Torque Estimation Accuracy: The method relies on accurate torque estimation from internal motor currents. This estimation can be affected by motor calibration, sensor noise, or thermal drift, potentially degrading performance, especially in prolonged or high-load tasks.
- Scalability to Other Modalities: It's unclear how scalable the framework is when extending to other physical modalities (like tactile or temperature sensing) under limited token budgets in transformer architectures.
- Future Directions: The authors suggest future work should focus on evaluating robustness in more diverse, real-world scenarios and exploring the alignment and integration of richer multimodal signals.