Short answer
Force/torque (F/T) data records the forces and moments a robot applies or feels, typically from a wrist-mounted six-axis sensor, joint torque estimates or a force plate. It is often the most accessible contact channel for tasks such as insertion, wiping, tightening and carrying. A buyer should specify sensor placement, frames, gravity compensation, sampling rate and synchronization, and should not expect F/T to describe where on the hand contact occurred.
The buyer's problem
Video shows that a part went into a fixture; it does not show how hard it was pushed or when it jammed. Policies for contact-rich work can benefit from a force channel, but force data is tied to a specific sensor and arm and is easy to misuse, so check what exists for your hardware before assuming it can be bought. A buyer needs to know what the numbers mean before paying for them.
What published data includes
RH20T contains more than 110,000 contact-rich real-world manipulation sequences, each with visual, force, audio and action data, across multiple robots and camera viewpoints, and every robot sequence is paired with a human demonstration video and a language description (Fang et al., 2023). It shows the pattern a buyer may want to copy: force recorded together with actions and views rather than as a standalone file.
By contrast, a portable hand-tracking rig such as DexCap has no force sensing (Wang et al., 2024), so human-demonstration systems based on pose alone carry no force channel. Force enters a dataset only when someone instruments the robot, the tool, the object or the floor. See motion capture data for humanoid motion for the force-plate option and tactile sensing data for distributed contact.
What an F/T channel does and does not tell you
| Signal | What it gives | What it hides |
|---|---|---|
| Wrist six-axis wrench | Net force and moment at the sensor | Where on the hand contact occurred; internal grip force |
| Joint torque or motor current | Effort per joint | Contact location; friction and model errors |
| Force plate | Ground reaction on the floor | Hand-object force |
| Tool-mounted sensor | Force at the tool | Differences between tools |
| Simulated force | Cheap and fully observable | Real friction, compliance and noise |
A wrist wrench includes the weight of everything beyond the sensor and inertial effects during motion. Without gravity and tool compensation, it is hard to separate contact force from payload. Ask whether compensation was applied, what tool and payload parameters were used and whether raw readings are kept.
Specifying the data
Put these parameters in the brief; each depends on the robot, task and policy.
- Sensor and placement. Model, range, resolution, mounting between flange and gripper, cable routing.
- Frame conventions. Sensor frame, tool frame and base frame, with transforms recorded as in calibration requirements.
- Bias and compensation. Zeroing procedure, gravity compensation, payload identification and drift checks.
- Sampling and filtering. Native rate, any low-pass filter, and the rate relative to the control loop; see action space and control frequency.
- Synchronization. Clock source and maximum offset to joint states and cameras; see sensor synchronization requirements.
- Overload events. Flags when the sensor saturated or the safety limit tripped.
- Task labels. Contact phases such as approach, contact, insertion, release, with who labeled them.
Illustrative F/T episode record
# Illustrative - synthetic values
episode_id: ep_000412
robot: {arm: "6-dof arm", gripper: "parallel jaw", control_hz: 20}
wrench_sensor:
id: ft_wrist_01
frame: sensor
native_rate_hz: 1000
stored_rate_hz: 500
filter: "low-pass, cutoff recorded in metadata"
bias_zero_at: episode_start
gravity_compensation: {applied: true, payload_kg: 0.85, params_file: payload_ep_000412.json}
streams: [wrench_raw, wrench_compensated, joint_pos, joint_torque, ee_pose, action, rgb_wrist, rgb_ext_1]
phase_labels: [approach, contact, insert, release]
outcome: {success: true, jam_events: 1}
Acceptance checks
- Bias check. In free motion without contact, compensated force stays near zero within a limit you set.
- Known-load check. Hang a reference mass and compare readings, at start and end of a session.
- Saturation count. Number and share of samples at sensor limits.
- Timing check. Contact onset aligns across force, joint states and video within tolerance.
- Phase label audit. A sample of labeled contact phases is reviewed against signal and video.
- Episode completeness. Every episode has all streams; missing streams are flagged, not zero-filled.
These plug into acceptance testing for deliveries.
Where F/T data comes from
Robot teleoperation setups can record wrist F/T alongside actions, which links to teleoperation data collection requirements. Some teleoperation interfaces give the operator no force feedback, so the recorded demonstrations may encode contact without the operator feeling it; ask how this affected behavior. Specialist sensing providers and research labs can run instrumented capture on your arm or an identical stack. Instrumented tools and fixtures can add force data to human demonstrations where the force is applied to a tool, and the device itself must be calibrated.
Business footage and document archives typically do not contain force data. Machine logs from industrial equipment may include torque or current, but those relate to the machine, not a human hand or a humanoid.
Reading force traces: what to ask the supplier to show
A buyer reviewing a sample should ask for plots, not summaries. Useful views are the wrench components against time with phase labels overlaid, the same trace beside synchronized wrist-camera frames, and a histogram of force magnitude per task. These show whether contact phases are visible, whether compensation left a visible offset, and whether the operator pressed so hard that the sensor clipped. Where teleoperators gave no haptic feedback, forces may show hesitant or excessive pushes that a trained policy could copy. Ask for the force limits enforced by the robot's safety layer, since a clipped trace may reflect a protective stop rather than the task.
What F/T data cannot support
F/T data cannot show object identity, grasp location or task goals on its own. It is tied to one arm, sensor and payload, so transfer to another embodiment is unproven without testing. Humanoid hands with many degrees of freedom need more than a wrist sensor to describe finger forces. Force alone cannot drive a policy: it needs paired observations and actions. And the quality of any force channel depends on rarely visible details: bias, cable strain, thermal drift. Plan for additional preparation: resampling, filtering decisions, compensation and episode QA.
Questions to ask suppliers
- Which F/T sensor, range and resolution, and where is it mounted?
- Are raw and compensated signals both delivered?
- How were bias, gravity and payload handled, and are parameters stored per episode?
- What sampling and filtering were applied before storage?
- How is the force stream synchronized with joint states and video?
- Were operators given force feedback during teleoperation?
- How many samples saturated the sensor, and are overload events flagged?
- Are contact phases labeled, and by whom?
- Can you replicate the sensor and arm stack for a commissioned collection?
- What license terms cover commercial training and derived models?
Discuss your requirements
If your tasks depend on contact, describe the forces, tasks and hardware to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates license and delivery; if no existing asset fits, SourceX can scope a commissioned collection. A request does not guarantee a match. Start at the humanoid training data guide or the procurement hub.
Discuss your humanoid training-data requirements with SourceX
Sources
- RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — RH20T provides force data per contact-rich sequence alongside vision and actions.
- DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation — DexCap hand mocap has no force sensing.