Chapter 8: Contact Learning — Online Improvement Through Rich Sensing

Overview

The hard part of manufacturing manipulation often begins after the camera has already seen a plausible scene. Slip, over-force, insertion jams, surface damage, compliant-material deformation, and sensor drift are failures that a vision-only policy sees late or not at all. DIGIT and later Digit-style fingertips show that compact tactile sensing can be a real channel for robot-hand learning ^[1], ^[2]. ReSkin and AnySkin foreground a different requirement: tactile skin must be replaceable and maintainable in the field ^[3], ^[4].

The point is not that adding touch always improves a policy. Manufacturers must design which failure is visible through which sensor, how often that sensor needs calibration, and which quality trace verifies a policy update. AnyTouch, ForceVLA, Tactile-VLA, and Tactile-Language-Action work show tactile and force signals entering representation learning and VLA-style stacks ^[6], ^[8], ^[9], ^[10]. The governing thesis of this chapter is stricter: touch and force become useful production evidence only when measurement, timestamping, calibration, maintenance, intervention, and replay remain connected from demonstration to release.

After reading this chapter... - Separate contact-rich manufacturing failures from ordinary vision failures. - Explain how DIGIT, ReSkin/AnySkin, CoinFT, AnyTouch, ForceVLA, and Tactile-VLA operate at different layers. - Design sensor drift, replacement, calibration, and online improvement together with policy architecture. - Define the force/tactile replay and QA traces required for a contact-rich PoC. - Distinguish measured force from force inferred from video, current, deformation, or a learned model. - Evaluate a modality gain together with sensing lifecycle cost and the limits of factory validation.

Contact Failures and Required Observations

The table treats touch and force as a failure taxonomy rather than a single technology. Some cells need only wrist force/torque, some need fingertip tactile images, and some need maintainable replaceable skin more than higher resolution.

Why Vision Alone Under-Specifies Contact

A camera observes reflected light. It can estimate pose, geometry, motion, and sometimes deformation, but many mechanically distinct states project to nearly the same image. A peg touching the left wall of a hole and a peg floating a millimeter away may look identical under occlusion. A gripper may hold an object securely or be one vibration away from slip while the visible pose remains unchanged. A cloth may look smooth while local tension is rising. This is not merely a shortage of pixels. It is an observability problem: the variables that determine the next safe action are not uniquely determined by the visual frame.

History makes the present frontier easier to read. Classical impedance and hybrid force-position control made contact force a controlled quantity, while analytical grasping formalized friction cones and force closure ^[25], ^[26]. Vision-based tactile devices later converted local deformation into dense images; DIGIT made this style of fingertip compact and comparatively accessible ^[1]. Replaceable magnetic skins such as ReSkin and AnySkin then shifted attention from resolution alone to reproducible replacement ^[3], ^[4]. Current tactile and force-aware foundation-policy work builds on those layers; it does not make the controller, calibration fixture, or maintenance process disappear.

Touch and force should therefore be logged as synchronized policy observations and outcome evidence, not attached only to a low-level controller. In the reported ForceVLA, TLA, and ForceVLA2 setups, contact signals expose state that visual-language inputs alone do not specify ^[8], ^[10], ^[31]. That evidence is bounded to the cited robots, sensors, tasks, baselines, and denominators. It does not prove universal transfer, a lower total data bill, or production readiness. The production implication is narrower and more actionable: if contact changes the action, the exact signal and its quality context must survive in the episode record.

This distinction also limits what a video world model can supply. A model may infer a plausible action or even a plausible force profile from visible motion. It cannot retroactively measure an unobserved six-axis wrench, contact pressure field, material damage, or calibrated torque. Video-derived labels can prioritize demonstrations and bootstrap a policy, but actual force and tactile channels remain necessary wherever release depends on overload, surface damage, seating, leak, or slip thresholds.

A Sensor Is Hardware and Data Policy

DIGIT made low-cost compact high-resolution tactile fingertips credible ^[1]. Later Digitizing Touch work made multimodal fingertips more clearly part of a robot-learning stack ^[2]. ReSkin and AnySkin address another issue: if tactile sensing is to leave the lab, skin must wear, be replaced, be recalibrated, and still preserve data lineage ^[3], ^[4].

Small 6-axis force/torque sensors such as CoinFT bring force closer to the finger or end-effector ^[5]. The manufacturer's decision is not only the sensor type. Sensor ID, calibration version, mounting torque, replacement date, cleaning cycle, and temperature range need to appear in the episode log. Otherwise a model regression cannot be separated from sensor drift.

Tactile Representation Is Translation Across Sensors

AnyTouch matters because it tries to unify static and dynamic tactile signals across multiple visuotactile sensors ^[6]. Factory sensing is rarely uniform. One line may use DIGIT fingertips, another replaceable skin, and another only wrist F/T plus cameras. If tactile representation is sensor-specific, the data flywheel breaks at every line boundary.

Figure 8.1. AnyTouch Figure 1 shows how TacQuad collects finely and coarsely aligned contacts at the same objects and locations with GelSight Mini, DIGIT, DuraGel, and Tac3D. It depicts a dataset of differing measured sensor outputs, not a performance result for the learned shared embedding. Source: Feng et al. 2025, arXiv:2502.12191 Fig. 1.

Sparsh, AnyTouch, and Sensor-Invariant Tactile Representation represent a progression from self-supervised features toward broader sensor transfer ^[27], ^[6], ^[28]. They are best understood as attempts to learn a common coordinate system, not proof that sensors have become interchangeable. A feature useful for texture classification can fail for shear-sensitive slip recovery. A representation aligned across two optical fingertips may not transfer to a magnetic skin, a wrist force/torque transducer, or a different frame rate. Pretraining benchmarks also do not reproduce contamination, pad wear, adhesive aging, remounting, or a changed cleaning agent.

Cross-sensor tactile representations are useful only to the extent that they survive changes in sensor geometry, temporal response, replacement, and calibration in the reported setup ^[27], ^[6], ^[28]. This is a qualified claim, not a promise of universal sensor invariance. A plant should verify it with leave-one-sensor-out tests, old-versus-new pad splits, and calibration-revision splits before pooling episodes across cells.

Synchronization is part of representation. RGB may arrive at 30 Hz, a tactile camera at another rate, a force/torque sensor at hundreds or thousands of samples per second, and joint states on the control clock. Naively concatenating the nearest timestamps can pair pre-contact video with post-impact force. The policy then learns a misleading causal order. Every episode should preserve hardware timestamps, clock source, interpolation rule, dropped packets, filtering delay, and the transform from sensor frame to tool frame. For a rapid jam or slip, milliseconds can decide whether the model learns detection or merely learns the aftermath.

Decision Walkthrough: Press-Fit Insertion

Press-fit insertion is a canonical contact-rich learning problem. To the camera, the peg and hole may look aligned, the wrist may descend smoothly, and the part may appear seated. The real failure occurs inside contact. Vision often sees too late whether the peg caught the chamfer, hit a burr, flexed the housing, or damaged the product through excessive force.

The necessary signal changes by stage. RGB and pose can be enough for approach and coarse alignment. First contact needs axial force and torque spikes. Micro-search needs force gradients and tiny retreat events. Final seating needs force drop, insertion depth, and inspection result together. Small force/torque sensors such as ^[5] can move this signal closer to the end-effector, while tactile fingertips can observe contact patch and slip more directly ^[1], ^[2].

Policy design also changes by stage. The high-level policy chooses the next approach pose. The local controller enforces force thresholds. A recovery policy tries small retreat and angle correction when a jam is detected. The QA gate reads both post-seating inspection and force trace. If press-fit is logged as a single success/fail episode, the team cannot tell which stage failed.

Sensor Lifecycle Is Data Lifecycle

Contact sensors behave more like consumables than fixed cameras. The replaceability emphasized by ReSkin and AnySkin is central to field deployment ^[3], ^[4]. In a factory, skin becomes contaminated, worn, cleaned, and sometimes replaced during a shift. When the sensor changes, the raw signal distribution changes. If that change is not logged, model updates and sensor drift become indistinguishable.

Sensor drift, contamination, damage, and replacement can invalidate a policy even when its model weights are unchanged. Modular GelSight-family designs make optics, illumination, gel, and housing trade-offs more systematic, but their modules still require calibration; reviews of vision-based tactile force measurement similarly show that marker tracking, photometric reconstruction, finite-element inversion, and learned force decoding carry different accuracy, range, and bandwidth assumptions ^[32], ^[33]. The production claim is therefore conditional: replacement improves serviceability only if the new unit can be related to the old unit through acceptance tests and logged calibration.

A useful sensor-lifecycle log includes sensor serial, mounting location, calibration file, cleaning event, replacement reason, hours of use, and temperature range. For tactile image sensors, gel condition and camera exposure matter. For wrist F/T, bias reset and tool-change events matter. For fingertip sensors, pad wear and contact surface matter.

This lifecycle must be tied to ML dataset version. Suppose policy v12 improves slip replay but deployed slip failures return. The cause may be model regression, a new sensor batch, or a cleaning-solvent change. Without sensor metadata, contact-rich learning loses reproducibility.

A practical acceptance sequence has four layers. First, a fixture applies repeatable normal and shear loads before installation. Second, the mounted sensor repeats task-relevant contacts with the production tool. Third, an overlap run compares the old and new sensor on the same replay set. Fourth, a restricted canary checks quality outcomes under the actual material and cleaning process. A single zeroing operation cannot replace this sequence because it tests offset, not dynamic response, spatial sensitivity, hysteresis, or the policy's learned dependence on the old signal.

Boundary Between Tactile Pretraining and Plant Fine-Tuning

Observability boundaries for contact signals — chapter synthesis

Unified tactile representation, as in AnyTouch, can improve transfer across sensors ^[6]. It does not eliminate plant fine-tuning. A factory has specific material, surface finish, lubrication, dust, temperature, and worker handling. Pretraining gives contact vocabulary; local fine-tuning learns how that vocabulary maps to product quality.

Manufacturers can divide tactile pretraining data into three layers. The first is sensor-general representation data: slip, press, roll, and tap across multiple sensors. The second is task-family data: insertion, wiping, cable handling, and food handling. The third is plant-specific data: replay built from the actual fixture, material, and inspection criteria.

Mixing these layers creates confusion. A strong sensor-general representation does not guarantee fewer plant-specific over-force events. Task-family data can help while still failing under a different cleaning protocol. Release gates should ultimately rely on plant-specific replay. Pretraining accelerates the start; fine-tuning and validation remain the plant's responsibility.

Operating Board for Force/Touch Online Improvement

A contact-rich policy needs an operating board of its own. This board does not track model metrics alone. It tracks force-limit violations, tactile dropout, sensor recalibration, pad replacement, operator recovery, scrap code, and rework cost. If slip success improves while pad replacement frequency doubles, the update may be operationally bad. If insertion success rises while over-force near misses rise, release should stop.

The board reviews candidate updates in three stages. Offline replay checks contact-failure buckets. Shadow mode reads live sensor streams but does not execute action, allowing anomaly predictions to be compared. Restricted release runs on a low-risk SKU or one cell. These stages reveal whether force/touch channels improve quality without creating hidden maintenance cost.

Without this structure, online improvement turns into faith that more data will help. In contact-rich tasks, more data also brings more sensor drift and maintenance variation. The operating board separates those effects.

ForceVLA and Tactile-VLA Reduce a VLA Blind Spot

VLAs connect language instructions and visual context to action. In contact-rich work, however, instruction and RGB miss the central state. ForceVLA uses force-aware routing for contact-rich manipulation ^[8]. Tactile-VLA introduces tactile observations as a policy channel and links physical knowledge to tactile generalization ^[9].

The recent methods should be compared by what is actually observed, not by whether “force” or “touch” appears in the title.

The measured-versus-inferred distinction matters most around thresholds. A learned estimator may rank “more force” and “less force” well enough to select demonstrations. A release rule such as “never exceed the fixture's validated wrench envelope” requires calibrated measurement and an independent guard. Likewise, a dense tactile image can reveal contact geometry without being a traceable six-axis force measurement unless the sensor and decoder were calibrated for that purpose ^[33].

Force-aware policies improve outcomes in several reported contact-rich setups, but present evidence remains bounded by sensor placement, robot, task family, controller, and evaluation protocol ^[8], ^[9], ^[31]. Modality benefit must be separated from a slower or safer controller, extra demonstrations, and additional instrumentation. An ablation that removes force while leaving the controller and data fixed is more informative than a comparison that changes the whole stack.

Figure 8.2. ForceVLA Figure 3 projects sensor-measured force f_t into force tokens, fuses them with visual and language tokens through a four-expert FVLMoE router, and sends the result to an action expert. After execution, f_{t+1} is measured again rather than inferred from the image. Source: Yu et al. 2025, arXiv:2505.22159 Fig. 3.

This direction is important, but attaching sensor tokens to a VLA is not enough. Force and tactile tokens create inference cost, synchronization issues, and missing-sensor behavior. Release gates should specify whether the policy stops, falls back to vision-only, or switches to guarded motion when a sensor drops out.

Figure 8.3. Tactile-VLA Figure 1 illustrates (a) tactile-aware instruction following, (b) inferred grip-force choice for an unseen object using learned tactile-relevant common sense, and (c) reasoning-driven pressure correction after a wiping failure. It is a capability teaser, not an architecture diagram of raw tactile measurement or sensor calibration. Source: Huang et al. 2025, arXiv:2507.09160 Fig. 1.

Sensor Procurement Should Be Written as an ML Requirement

Buying force or tactile sensors from a hardware spec sheet is not enough. Repeatability, bandwidth, resolution, and overload tolerance matter, but ML requirements often matter more: calibration export, timestamp synchronization, raw trace access, and replacement reproducibility. A sensor can produce good measurements and still be poor training data if it is not aligned to episodes.

Acceptance tests should be task-specific. In an insertion cell, test axial force spikes and torque signatures. In a wiping cell, test pressure distribution and surface contamination. In cable handling, test slip and tension. In food handling, test compliance and post-cleaning drift. The same sensor can have different useful ranges and failure signatures across tasks.

Responsibility boundaries are also needed when the sensor vendor and policy vendor differ. If calibration changes, does the model vendor adapt the policy, or does the manufacturer manage the normalization layer? If raw tactile images create privacy or process-IP issues, who redacts them? If these questions are not written at procurement time, sensor data often fails to become policy-improvement data at deployment time.

A contact-rich PoC should therefore specify robot, hand, sensor, and policy separately while unifying timestamp and episode lineage. The sensor is not just the policy's eye or skin; it is part of the quality trace.

Shadow-Mode Tactile Experiment

Putting a contact sensor directly into closed-loop control can be risky. A lower-risk path is shadow mode. The robot continues to move under the existing vision policy or guarded controller, while a tactile/force model predicts slip, jam, or over-force in real time. The prediction does not affect action. Later it is compared with QA labels and operator notes. This tests whether the new channel adds information before it receives authority.

A shadow experiment should answer three questions. First, how much earlier does the sensor event appear than inspection failure? Second, are false positives frequent enough to damage cycle time or operator trust? Third, does the event appear only for a specific SKU, pad-wear state, cleaning state, or fixture revision? Until these questions are answered, a tactile token inside a policy may be improvement or noise.

After shadow mode, the cell can move to supervised assist. In that phase, the sensor model proposes stop, retreat, or regrasp, but a human or existing controller approves the action. Only after that should restricted autonomy allow sensor-driven recovery on a low-risk SKU. Contact-rich online improvement should move control authority gradually.

This staged approach also protects evaluation. If a tactile model is introduced directly into autonomy and success improves, the team may not know whether the improvement came from better sensing, more conservative motion, or changed operator behavior. Shadow and supervised stages separate prediction quality from control authority. That separation is essential when a sensor channel is new and maintenance behavior is still unknown.

The same staged logic applies when replacing sensors. A new tactile skin or force sensor batch should first run in shadow against the old channel. If its events disagree, the disagreement becomes a calibration task, not a policy failure. Only after the replacement sensor explains the same contact replay should it be allowed to drive recovery.

Shadow mode also makes operator training easier. Maintenance staff can see what the sensor would have flagged before the robot is allowed to act on it. Quality engineers can compare sensor events with inspection defects. Operators can report cases where the model flags a harmless contact pattern. Those comments become labels for the next calibration pass.

The useful output is not only a better classifier. It is an operations vocabulary. A slip warning, a jam warning, and an over-force warning should each have an owner, a response, and a review cadence. When a tactile model produces a new warning type, the cell should not automatically add another alarm. The team should decide whether the warning changes action, maintenance, QA sampling, or only future training data.

That discipline keeps contact sensing from becoming alarm noise. Operators will ignore a sensor that only adds interruptions. They will trust it when each warning has a clear physical interpretation and a visible path into either recovery behavior or maintenance action.

It also gives QA a way to audit model changes without becoming tactile-sensor specialists.

That audit path is what makes touch operationally credible.

Online Improvement Starts With Contact Replay

Online correction and replay should prioritize contact transitions and near-failures that vision-only success labels miss. Failures must become contact replay: slip replay needs pad-wear and contamination splits; insertion-jam replay needs approach angle, axial force, and retreat timing; over-force replay must connect the force trace to cosmetic or dimensional inspection. UMI-FT-style demonstrations broaden compliant data capture ^[11] [#36], while human-in-the-loop reinforcement learning shows how interventions can be combined with demonstrations, a learned reward classifier, off-policy updates, and engineered reset/control choices ^[34]. Neither source establishes unattended, monotonic improvement on a production line.

The operating loop has three steps. First, a failure detector catches force or tactile anomalies. Second, a human or guarded controller performs recovery and records the correction. Third, the next model update must pass both contact replay and a production canary before release. Without that loop, a tactile sensor becomes dashboard decoration rather than policy evidence.

Interventions are especially valuable when they record the boundary between acceptable and unacceptable contact. A stop just before a scratch, a retreat after a torque spike, or a regrasp before incipient slip provides more information than another routine success. The log should retain the pre-intervention window, the operator or guard that intervened, the proposed action, the executed correction, the sensor-health state, and the final quality result. Treating every intervention as a failed episode discards this structure; treating it as a privileged correction preserves a counterfactual training signal.

Improvement Inside Safety and Quality Envelopes

Contact-sensor evidence ladder — chapter synthesis

“Online” should describe the arrival of evidence, not unrestricted online exploration. A contact policy can improve only inside an envelope whose hard limits remain outside the learned optimizer. The safety side includes force, torque, speed, workspace, collision, temperature, and protective-stop constraints. The quality side includes scratch class, seating depth, leak result, deformation, torque signature, and downstream inspection. A candidate policy is allowed to optimize cycle time or success only after satisfying both.

The envelope should be implemented at more than one layer. A fast controller or safety PLC enforces physical limits. The learned policy chooses approach, force target, or recovery within those limits. A monitor checks synchronization, sensor health, and out-of-distribution conditions. The release process checks replay strata and quality outcomes. This decomposition prevents a foundation policy from becoming the sole interpreter of its own safety evidence.

For each update, the team should publish a small release card: changed data slices; sensor and controller versions; replay results by failure type, product variant, and sensor age; confidence intervals or trial counts; new near misses; maintenance burden; rollback trigger; and the person authorized to expand the canary. Success rate alone hides exactly the evidence a contact-rich cell needs.

Offline replay also needs “do no harm” strata. A change targeting slip recovery must not raise over-force or cycle-time tails. A change targeting insertion must not pass by becoming so conservative that throughput collapses. A new tactile representation must retain performance after sensor replacement, not only on randomly held-out episodes from the same device. These checks turn online learning from continual weight change into controlled evidence accumulation.

Maintenance and Drift Belong Inside the Model Story

Tactile and force sensors age with the production equipment. Skins wear, gel becomes contaminated, cables loosen, mounting changes, and calibration files update. If those changes are not logged, model regression is hard to diagnose. When the same policy is deployed across cells, differences in sensor revision and cleaning protocol become hidden domain shifts.

Sensor maintenance has to be part of MLops. Model version, dataset version, and controller version should be tied to sensor calibration version and replacement events. Rolling back a tactile-rich policy may require reverting thresholds, force guards, and calibration jobs, not only weights.

Manufacturing Cell Checkpoint

First, rewrite task failures in tactile and force terms. An "insertion failure" may be an alignment error, axial force spike, burr contact, or soft-part deformation. Second, choose the minimum sensor set needed for that failure. Do not force fingertip tactile into a task where wrist F/T is sufficient, and do not push fingertip-contact tasks through a vision-only stack.

Third, build a contact replay set: slip, over-force, jam, deformation, sensor dropout, and calibration drift. Fourth, confirm that the vendor or internal team can export raw force/tactile traces and calibration metadata. Without owning that data, the manufacturer cannot own the cause of online improvement.

Open Questions and Failure Modes

First, tactile data is harder to collect and normalize than vision data, and every sensor family has its own representation. Second, force/tactile tokens give a VLA more state while increasing inference and synchronization burden. Third, even strong sensors cannot teach the relation between over-force and cosmetic damage without QA labels. Fourth, missing maintenance logs make drift look like policy failure and policy failure look like drift.

The strongest counterevidence is the narrowness of current validation. ForceVLA and ForceVLA2 report results on five tasks in their respective setups; Tactile-VLA uses a small task set and a sensor-specific encoding; UMI-FT reports three single-arm tasks; DexForce depends on two force/torque sensors, kinesthetic teaching, and per-task compliance choices ^[8], ^[31], ^[9], ^[11], ^[29]. These are meaningful method demonstrations, not independent factory-scale lifecycle studies. Public evidence does not yet establish performance across months of wear, multiple sensor lots, cleaning regimes, heterogeneous cells, or third-party maintenance.

There is also a denominator problem. A modality can improve average task success while increasing false stops, recalibration labor, pad consumption, or tail-force events. Those costs are usually absent from research benchmarks. Factory validation should therefore report not only completed trials but resets, excluded trials, interventions, sensor replacements, maintenance minutes, damaged parts, and post-inspection failures. Without those denominators, a modality gain cannot be converted into an operating case.

Finally, richer sensing can create false confidence. A tactile image is not automatically a force measurement; a learned force estimate is not automatically calibrated; a calibrated sensor is not automatically synchronized; synchronized data is not automatically causal; and a successful benchmark is not automatically safe deployment. Each arrow requires its own evidence.

Figure 8.4. Online contact-learning gate from event to quality inspection. This diagram is not a performance claim; it summarizes the operational chain that carries capability evidence through replay and quality decisions into release authority. Source: author-created SVG.

What to Learn Next

Chapter 9 reads these technologies through company strategy. Foundation-model labs, production-flywheel companies, and hand/hardware co-design companies all talk about data scale, but they own different pieces. The contact-evidence test developed here becomes a due-diligence question there: who owns the raw sensor trace, calibration history, intervention log, quality label, and authority to roll back? The next chapter uses PI, Generalist, Skild, and Figure to ask who owns the model, data, hardware, and deployment evidence.

References

1. Lambeta, Mike (2020). DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand Manipulation. arXiv.
2. Lambeta, Mike (2024). Digitizing Touch with an Artificial Multimodal Fingertip. arXiv.
3. Bhirangi, Raunaq (2021). ReSkin: Versatile, Replaceable, Lasting Tactile Skins. arXiv.
4. Bhirangi, Raunaq (2024). AnySkin: Plug-and-play Skin Sensing for Robotic Touch. arXiv.
5. Choi, Hojung (2025). CoinFT: A Coin-Sized, Capacitive 6-Axis Force Torque Sensor for Robotic Applications. arXiv.
6. Feng, Ruoxuan (2025). AnyTouch: Learning Unified Static-Dynamic Representation across Multiple Visuo-Tactile Sensors. arXiv.
7. Yang, Fengyu (2023). Touch and Go: Learning from Human-Collected Vision and Touch. arXiv.
8. Yu, Jiawen (2025). ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation. arXiv. #1
9. Huang, Jialei (2025). Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization. arXiv.
10. Hao, Peng et al. (2025). TLA: Tactile-Language-Action Model for Contact-Rich Manipulation. arXiv.
11. Choi, Hojung (2026). In-the-Wild Compliant Manipulation with UMI-FT. arXiv. #36
12. Xu, Mengda (2025). DexUMI: Using Human Hand as the Universal Manipulation Interface for Dexterous Manipulation. arXiv.
13. Fang, Hao-Shu (2025). DEXOP: A Device for Robotic Transfer of Dexterous Human Manipulation. arXiv.
14. Si, Zilin (2025). ExoStart: Efficient learning for dexterous manipulation with sensorized exoskeleton demonstrations. arXiv.
15. Shaw, Kenneth (2023). LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning. arXiv.
16. Figure AI (2026). Figure 03 + Helix 02: General-Purpose Humanoid System. Company product page.
17. Zhao, Tony Z. (2023). Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. arXiv.
18. Chi, Cheng (2023). Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. arXiv.
19. Black, Kevin (2024). pi0: A Vision-Language-Action Flow Model for General Robot Control. arXiv.
20. Kim, Moo Jin (2024). OpenVLA: An Open-Source Vision-Language-Action Model. arXiv.
21. Khazatsky, Alexander (2024). DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset. arXiv.
22. O'Neill, Abby (2023). Open X-Embodiment: Robotic Learning Datasets and RT-X Models. arXiv.
23. AgiBot-World-Contributors et al. (2025). AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems. arXiv.
24. Toyota Research Institute (2024). Large Behavior Models for Robot Manipulation. Company technical post.
25. Hogan, Neville (1985). Impedance Control: An Approach to Manipulation: Part I—Theory. Journal of Dynamic Systems, Measurement, and Control.
26. Khatib, Oussama (1987). A Unified Approach for Motion and Force Control of Robot Manipulators: The Operational Space Formulation. IEEE Journal on Robotics and Automation.
27. Higuera, Carolina et al. (2024). Sparsh: Self-Supervised Touch Representations for Vision-Based Tactile Sensing. CoRL.
28. Gupta, Harsh et al. (2025). Sensor-Invariant Tactile Representation. ICLR.
29. Chen, Claire et al. (2025). DexForce: Extracting Force-informed Actions from Kinesthetic Demonstrations for Dexterous Manipulation. IEEE Robotics and Automation Letters. #3
30. Adeniji, Ademi et al. (2025). Feel the Force: Contact-Driven Learning from Humans. arXiv.
31. Li, Yang et al. (2026). ForceVLA2: Unleashing Hybrid Force-Position Control with Force Awareness for Contact-Rich Manipulation. arXiv.
32. Agarwal, Arpit et al. (2025). A Modularized Design Approach for GelSight Family of Vision-based Tactile Sensors. The International Journal of Robotics Research.
33. Fang, Bin et al. (2025). Force Measurement Technology of Vision-Based Tactile Sensor. Advanced Intelligent Systems.
34. Luo, Jianlan et al. (2025). Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning. Science Robotics.
35. Fengyu Yang et al. (2024). Binding Touch to Everything: Learning Unified Multimodal Tactile Representations. CVPR 2024.
36. Sandra Q. Liu et al. (2024). A Passively Bendable, Compliant Tactile Palm with RObotic Modular Endoskeleton Optical (ROMEO) Fingers. IEEE International Conference on Robotics and Automation (ICRA).
37. Jessica Yin et al. (2024). Learning In-Hand Translation Using Tactile Skin With Shear and Normal Force Sensing. arXiv preprint arXiv:2407.07885.
38. Branden Romero et al. (2024). EyeSight Hand: Design of a Fully-Actuated Dexterous Robot Hand with Integrated Vision-Based Tactile Sensors and Compliant Actuation. IROS 2024.
39. Binghao Huang et al. (2024). 3D-ViTac: Learning Fine-Grained Manipulation with Visuo-Tactile Sensing. CoRL 2024.
40. Han Zhang et al. (2025). DOGlove: Dexterous Manipulation with a Low-Cost Open-Source Haptic Force Feedback Glove. RSS 2025.
41. Akash Sharma et al. (2025). Self-supervised perception for tactile skin covered dexterous hands. arXiv preprint arXiv:2505.11420.