FACTR 2: Learning External Force Sensing for Commodity Robot Arms Improves Policy Learning
# FACTR 2: Force Sensing Without Expensive Sensors
Researchers have developed a software method called Neural External Torque Estimation (NEXT) that allows standard robot arms to detect contact forces using only their existing joint motors—eliminating the need for specialized force sensors. The system learns to estimate external forces by analyzing motor behavior during brief periods of unloaded movement, then applies this learned model during actual manipulation tasks. Training requires approximately one minute of computation using about ten minutes of free-motion data collected from the robot itself.
For automation integrators and logistics operators, this addresses a significant cost barrier in contact-heavy operations. Many industrial robot arms lack force feedback capabilities because dedicated force sensors add substantial expense and complexity to system architecture. A software-only approach that runs on existing hardware could expand force-sensitive manipulation to a broader range of deployed systems, particularly in bin picking, assembly, and deformable material handling where contact detection currently requires either expensive sensor additions or alternative workaround solutions.
The practical implication worth noting is that effectiveness appears tied to the specific robot model used during training. Whether NEXT estimates generalize reliably across different arm designs, payload configurations, or significantly worn mechanical systems remains an open operational question. Implementation in production environments would likely require baseline testing on each specific hardware platform before deployment in critical tasks.