Researchers have developed a system that lets low-cost robot arms sense external forces using only motor current and 10 minutes of free-space movement. The method, NEXT, generates torque estimates comparable to dedicated sensors, enabling force-feedback teleoperation and a training technique called FIRST that boosts success rates on contact-rich tasks.
What the Researchers Built
The team introduced FACTR 2, a two-component system that brings force awareness to commodity robot arms — without requiring expensive force-torque sensors. The first component is NEXT (Neural External Torque estimation), a data-driven method that estimates external joint torque by learning the robot's free-space dynamics using only a dozen minutes of motion data. The second component is FIRST (Force Informed Re-Sampling Training), a behavior cloning technique that leverages NEXT’s torque signals to identify and emphasize the most critical parts of demonstrations: the moments just before contact and during contact itself.
Previous approaches either relied on dedicated joint torque sensors (found only on high-end arms like the Franka Panda, costing $30,000) or used model-based disturbance observers that require detailed system identification. NEXT learns a neural network from scratch, needing only 10 minutes of free-space trajectories and 1 minute of training. FIRST then uses the estimated torque to segment each demonstration into free-space, pre-contact, and contact phases, and up-samples the latter two to force the policy to learn from the transitions where failures typically occur.
The work was validated on five long-horizon, contact-rich tasks, including LEGO assembly, belt assembly, peg insertion, tool clean-up, and cap screwing. The approach generalizes across arms from the $2,500 AgileX Piper to the $30,000 Franka Panda.
Key Results
A user study with 20 participants compared force-feedback teleoperation on the Franka arm under five conditions: no feedback, disturbance observer, position-position feedback, FACTR Teleop using built-in torque sensors, and FACTR Teleop using NEXT’s estimates. Participants consistently rated NEXT-based feedback as easier to use, with scores near those of the sensorized version. On the low-cost Piper arm, NEXT also outperformed baseline feedback methods.
For policy learning, FIRST was evaluated against several baselines including standard ACT, ACT with raw torque input, FACTR (visual curriculum), and TA-VLA (auxiliary torque target). Across all five tasks, FIRST improved task progress rate by 10–25 percentage points compared to base policies, with the largest gains in tasks requiring precise alignment (e.g., NIST insertion) and controlled contact (e.g., cap screwing). An ablation study showed that moderate up-sampling (weight ≈ 5) works best; too aggressive up-sampling degrades performance by biasing the training distribution away from free-space motion.
The method was tested with two action representations (ACT and flow-matching) and showed consistent improvements, confirming that the gains come from the data re-sampling strategy rather than the policy architecture.
How It Works
NEXT estimates external joint torque by learning the robot's free-space inverse dynamics. First, a human or script moves each joint individually through its full range, then executes multi-joint Cartesian motions at varying speeds — totaling 10 minutes. This data is used to train a compact LSTM (recurrent neural network) that predicts the motor torque needed to follow a commanded trajectory in free space. At deployment, the measured motor torque (from motor current) is subtracted from the predicted free-space torque; the residual is the estimated external torque.
The key insight is that the learned model captures joint friction, inertia, and Coriolis effects without any robot-specific physics modeling. The LSTM runs at the control frequency (typically 100–500 Hz) and outputs a torque estimate for each joint.
FIRST then uses the NEXT torque signal to label each time step in a demonstration: if torque is below a low threshold, it's free-space; if it rises above a second threshold, it's contact; the interval between is pre-contact. During training, the policy batches are constructed by sampling from all phases equally, which effectively up-samples the shorter pre-contact and contact phases. This forces the policy to learn from the exact moments where policies usually fail — when the gripper is aligning with a target or applying force.
Why This Matters for Robotics
Force sensing has been a critical missing piece for affordable robot arms used in manufacturing, logistics, and research. Dedicated torque sensors add thousands of dollars per arm and require complex wiring and calibration. FACTR 2 eliminates that barrier: a $2,500 AgileX Piper can now feel contact as well as a $30,000 Franka.
For robot buyers and integrators, this means that contact-rich tasks — peg insertion, screw driving, snap-fitting, assembly — become feasible on cheaper hardware. Browse humanoid robots on Robot Overflow or used industrial robots for sale to see platforms that could benefit from such software-only force sensing. The approach also enables safer human-robot collaboration by detecting unexpected collisions without extra sensor investment.
The FIRST training technique is architecture-agnostic and could be applied to any imitation learning pipeline, making it a drop-in improvement for many existing robot learning workflows.
Limitations and Open Questions
NEXT assumes the robot's dynamics are consistent between free-space data collection and deployment — changes in payload or joint wear could degrade accuracy. The 10-minute data collection, while minimal, is still a manual step that might not scale to fleets of robots. The torque threshold for phase segmentation in FIRST is a hyperparameter that may need tuning per task. Additionally, the method has only been tested on rigid, quasi-static contact; highly dynamic tasks (e.g., hammering) or very deformable objects (e.g., cloth) could challenge the torque-based segmentation.
Finally, the user study was conducted in a lab setting with a limited number of participants; real-world teleoperation performance over long shifts is untested.
Frequently Asked Questions
Do I need any special hardware to use NEXT? No. NEXT uses only the motor current signals already available on most robot arms — no force-torque sensors, no external measurement devices.
How long does it take to set up NEXT on a new robot arm? About 10 minutes to collect free-space trajectories and 1 minute to train the neural network. The motion commands can be auto-generated or recorded once and replayed.
Can FIRST be used with any imitation learning policy? Yes. The authors tested FIRST with both ACT and flow-matching policies, and it improved performance in both cases. Any policy that trains on sequential demonstration data could adopt the re-sampling approach.
Does FACTR 2 work on both cheap and expensive arms? Yes. The method was validated on the low-cost AgileX Piper ($2,500) and YAM arms, as well as the high-end Franka Panda ($30,000). The force estimates matched dedicated sensor quality.
Conclusion
FACTR 2 shows that accurate force sensing is achievable on commodity robot arms with minimal data and no extra sensors. By combining a simple learned torque estimator with a smart training re-sampling technique, the system closes the gap between expensive sensorized robots and affordable ones. This could accelerate the adoption of force-aware manipulation in real-world production lines and service robotics.