FACTR 2: Learning External Force Sensing for Commodity Robot Arms Improves Policy Learning

Researchers introduce NEXT, a neural network method that estimates external joint torques on robot arms without dedicated force sensors, paired with FIRST, a training technique that improves policy learning by 17% across long-horizon tasks. This breakthrough enables cost-effective force-aware teleoperation and manipulation on commodity robots by leveraging only 10 minutes of free-motion calibration data.

The advancement of force sensing in robotics has long been constrained by hardware costs, limiting widespread adoption of force-feedback capabilities in commercial robot arms. NEXT addresses this fundamental constraint through a software-driven approach that infers joint torques from motor currents and kinematic data, requiring minimal calibration. This represents a significant accessibility improvement for roboticists and manufacturers seeking to deploy contact-rich manipulation tasks like insertion, assembly, and delicate object handling without expensive sensor upgrades.

The work builds on growing recognition that force awareness directly improves learning efficiency in behavioral cloning and imitation learning. By augmenting standard policy learning with Force-Informed Re-Sampling Training (FIRST), which strategically oversamples contact-critical phases of demonstrations, the method achieves measurable performance gains. The 17% improvement in task progress across five evaluated tasks demonstrates practical value beyond theoretical novelty.

This development has meaningful implications for the robotics industry's accessibility ladder. Off-the-shelf arms from manufacturers like UR, ABB, and KUKA already have the necessary motor encoders and current sensors; NEXT unlocks latent force-sensing capability without hardware modification. Companies developing autonomous manipulation pipelines, warehouse automation systems, and collaborative manufacturing solutions can now incorporate force feedback at reduced capital expense, potentially accelerating adoption timelines.

The open-source release and reproducible methodology invite rapid iteration from the research community. Future work will likely explore domain adaptation across different robot models and extended task complexity. The convergence of neural estimation and learning optimization represents a pattern where software increasingly compensates for hardware limitations, democratizing robot capabilities.

• →NEXT estimates external joint torques without dedicated sensors using only 10 minutes of calibration data and 1 minute of neural network training
• →FIRST training method improves policy learning performance by over 17% through strategic resampling of contact-rich demonstration phases
• →Approach requires no hardware modifications to existing commodity robot arms, only leveraging existing motor encoders and current sensors
• →Open-source implementation enables immediate adoption across industrial and research robotics applications
• →Method addresses a critical cost barrier that previously limited force-aware manipulation to expensive, specialized robotic systems