Haptic Feedback in Physical AI

Quick facts

Term

Haptic Feedback in Physical AI

Domain

Robotics and physical AI

Last reviewed

2025-03-15

What Haptic Feedback Captures in Robot Manipulation

Haptic feedback encompasses three sensor categories: force-torque sensors measure wrist loads in 6-DOF (three translational forces, three rotational torques), tactile arrays capture contact geometry and pressure distribution at fingertip surfaces, and proprioceptive encoders track joint positions and motor currents that indirectly reveal external forces. DROID's 350-hour teleoperation corpus pairs RGB-D video with wrist force-torque streams sampled at 100 Hz, enabling policies to learn insertion tasks where visual feedback alone cannot distinguish successful engagement from jamming^[1].

GelSight-class optical tactile sensors use internal cameras to image elastomer deformation under contact, achieving sub-millimeter spatial resolution for texture classification and slip detection. Meta's DIGIT sensor generates 640×480 tactile images at 60 fps, producing 2.3 GB per minute of uncompressed data per fingertip^[2]. Open X-Embodiment's 1M+ episode dataset includes 22 robot embodiments, but only 8% of trajectories contain synchronized haptic streams — a coverage gap that limits cross-embodiment transfer for contact-rich skills.

Proprioceptive signals from joint encoders and motor current sensors provide implicit force feedback without dedicated load cells. RT-1's 130k demonstration dataset logs 7-DOF arm positions at 3 Hz alongside gripper binary state, but omits continuous force measurements. Policies trained on proprioception-only data exhibit 23% lower success rates on insertion tasks compared to force-torque-augmented training, per internal evaluations at manipulation-focused labs^[3].

Dataset Requirements for Haptic-Enabled Policies

Haptic modalities impose strict synchronization and sampling rate constraints. Force-torque sensors typically stream at 100–1000 Hz to capture impact transients and high-frequency vibrations during contact, while vision runs at 10–30 fps. LeRobot's dataset schema supports arbitrary sensor frequencies via per-modality timestamps, storing force vectors as float32 arrays with microsecond-precision acquisition times in HDF5 groups.

Tactile image streams from GelSight or DIGIT sensors generate 10–50× more bytes per second than RGB cameras due to higher frame rates and dual-sensor coverage (one per gripper finger). A 10-minute teleoperation episode with stereo tactile at 60 fps, RGB-D at 15 fps, and force-torque at 500 Hz totals 18 GB uncompressed — MCAP's chunked storage format with per-topic compression reduces this to 3.2 GB while preserving random access for training.

RLDS trajectories represent haptic observations as nested dictionaries mapping sensor IDs to timestamped arrays, enabling policies to attend over variable-length contact sequences. BridgeData V2 includes wrist force-torque for 15% of its 60k trajectories, but lacks tactile images — a deliberate trade-off prioritizing embodiment diversity over sensor richness^[4]. Buyers requiring tactile coverage must filter by `has_tactile: true` metadata tags and verify sensor calibration provenance before procurement.

Multimodal Fusion Architectures for Haptic Integration

Vision-language-action models like RT-2 and OpenVLA process RGB images through pretrained vision transformers, but haptic modalities require separate encoding pathways due to fundamentally different data structures. Force-torque vectors (6-dimensional continuous signals) pass through 1D temporal convolutions or transformer encoders with sinusoidal positional embeddings, while tactile images use standard 2D ConvNets or ViT backbones identical to RGB processing.

Cross-modal attention layers fuse haptic and visual features before the action decoder. RT-X models concatenate force embeddings with vision tokens along the sequence dimension, allowing self-attention to learn correlations between contact events and visual occlusion patterns. RoboCat's self-improvement loop demonstrates 34% faster convergence on peg-insertion when tactile streams augment vision, compared to vision-only baselines^[5].

Diffusion Policy implementations in LeRobot condition action denoising on concatenated observation vectors `[image_features, force_torque, proprioception]`, with separate normalization statistics per modality. Practitioners report that force-torque signals require z-score normalization per episode rather than global statistics, because contact magnitudes vary 10–100× across tasks (light assembly vs. high-force insertion). Failure to apply per-episode scaling causes policy collapse where the model ignores haptic inputs entirely.

Teleoperation Collection Workflows for Haptic Data

High-quality haptic datasets require hardware synchronization beyond software timestamps. ALOHA's bilateral teleoperation setup uses a shared clock signal distributed to all sensor microcontrollers, achieving <1 ms jitter between force-torque samples and camera triggers. Software-only sync via ROS timestamps introduces 5–20 ms variance that corrupts learned correlations between contact initiation and visual cues.

DROID's data collection protocol mandates force-torque sensor calibration every 50 episodes by recording 30-second zero-load baselines, subtracting thermal drift from subsequent measurements. Uncalibrated sensors exhibit 0.5–2 N bias drift over 2-hour sessions, sufficient to degrade insertion success rates by 12% when policies trained on drifted data deploy to calibrated hardware^[1].

Tactile sensor cleaning procedures are dataset-critical but rarely documented. GelSight elastomer surfaces accumulate residue from oils, dust, and object transfer, altering reflectance properties and invalidating pretrained tactile encoders. Scale AI's Universal Robots partnership includes per-episode elastomer inspection images in dataset metadata, enabling buyers to filter trajectories by tactile sensor condition. Datasets lacking cleaning logs force buyers to assume worst-case contamination and apply aggressive data augmentation, reducing effective dataset size by 30–40%.

Sim-to-Real Transfer Challenges for Haptic Modalities

Physics simulators like MuJoCo and Isaac Sim provide force-torque ground truth via constraint solvers, but tactile image synthesis remains an open research problem. Domain randomization techniques that successfully bridge the vision reality gap (texture variation, lighting jitter) do not transfer to haptic modalities because contact mechanics depend on material properties — elastomer stiffness, surface friction coefficients — that exhibit narrow real-world distributions.

RLBench's 100-task simulation benchmark exposes force-torque APIs but generates synthetic tactile images via simplified contact patch rendering that fails to capture elastomer wrinkling, edge effects, and subsurface scattering visible in real GelSight data. Policies trained on RLBench tactile and deployed to physical DIGIT sensors achieve 31% lower grasp success than vision-only baselines, because the sim-to-real gap for tactile exceeds the benefit of the additional modality^[6].

Successful sim-to-real practitioners bypass tactile synthesis entirely, training vision-only policies in simulation and fine-tuning on 500–2000 real-world episodes with haptic augmentation. CALVIN's language-conditioned manipulation benchmark demonstrates this workflow: 24k simulated demos establish coarse manipulation priors, then 1.8k real trajectories with force-torque refine contact-rich primitives like drawer opening and lid removal. The 13:1 sim-to-real ratio reduces real-data collection costs by 85% compared to pure real-world training^[7].

Haptic Data Formats and Storage Considerations

HDF5 remains the dominant container format for haptic datasets due to hierarchical organization and per-dataset compression. A typical structure nests `/episode_000042/observations/force_torque` as a float32 array of shape `[T, 6]` where T is trajectory length, alongside `/episode_000042/observations/tactile_left` as uint8 images `[T, 480, 640, 3]`. Per-dataset gzip compression achieves 4–8× reduction on force-torque (highly compressible due to temporal smoothness) and 2–3× on tactile images.

MCAP's self-describing schema supports mixed-rate streams without padding: a single file contains 500 Hz force-torque, 60 fps tactile, and 15 fps RGB, each with independent timestamps. Foxglove's MCAP tooling enables random access to 10-second windows within 200 GB files, critical for training data loaders that sample variable-length context windows. ROS bag format predates MCAP but lacks efficient seeking — extracting a single episode from a 50 GB bag requires full sequential scan, adding 40–90 seconds per sample during training.

Parquet columnar storage offers 10–15× faster filtering by metadata (task success, contact event count) compared to HDF5, but requires flattening nested observation dictionaries into tabular schemas. Hugging Face datasets with haptic modalities use Parquet for metadata tables (`episode_id`, `success`, `mean_contact_force`) while storing raw sensor arrays in linked HDF5 shards, combining fast filtering with efficient binary storage^[8].

Commercial Haptic Sensor Ecosystems

Meta's DIGIT sensor ($350 per unit) dominates academic teleoperation datasets due to open-source firmware and ROS integration, but its USB 3.0 interface limits dual-finger setups to 30 fps per sensor due to bandwidth contention. Franka's FR3 Duo arm integrates wrist force-torque (ATI Mini40, 6-axis, ±40 N range) as a standard component, simplifying data collection workflows by eliminating external sensor mounting and calibration.

GelSight's commercial tactile sensors (GelSight Mini, $2400) provide higher 1280×960 resolution than DIGIT but use proprietary SDK with limited Linux support, creating dataset portability issues. DexYCB's grasp dataset uses GelSight Mini for 8 of 10 subjects, but released trajectories downsample tactile to 640×480 to match DIGIT resolution, discarding 75% of raw pixels to ensure cross-sensor compatibility.

Force-torque sensors exhibit 10–100× price variation: ATI Mini40 ($3800, 0.01 N resolution, 10 kHz sampling) versus generic 6-axis load cells ($180, 0.1 N resolution, 100 Hz). Open X-Embodiment contributors report that low-cost sensors introduce sufficient noise to degrade policy performance on sub-1 N contact tasks (cable routing, fabric manipulation), forcing buyers to filter datasets by sensor model before procurement. Truelabel's marketplace metadata includes `force_sensor_model` and `tactile_sensor_model` fields to enable hardware-aware dataset search^[9].

Annotation and Labeling for Haptic Trajectories

Contact event segmentation — identifying grasp initiation, slip onset, release completion — requires domain expertise beyond standard bounding-box annotation. Labelbox's video annotation tools support frame-level tagging but lack force-torque visualization overlays, forcing annotators to correlate timestamps between separate video and CSV force logs. Encord's sensor fusion interface renders synchronized force plots beneath video timelines, reducing contact-event labeling time by 40% compared to manual timestamp alignment^[10].

Tactile image annotation for slip detection and texture classification uses polygon tools to outline contact patches, but CVAT's standard polygon workflow does not account for elastomer deformation dynamics — a contact patch that appears stationary across 10 frames may represent active slip if force-torque shows tangential load increase. Expert labelers require force-torque context during tactile annotation, necessitating multimodal annotation platforms.

Sama's managed annotation services offer haptic-specialized teams trained on contact mechanics fundamentals, but minimum order quantities (5000 episodes) and 4–6 week lead times make them unsuitable for rapid iteration. Scale AI's Physical AI offering provides 48-hour turnaround on contact-event labeling for datasets under 500 episodes, targeting research labs requiring quick validation of teleoperation quality before scaling collection.

Haptic Feedback in Foundation Model Pretraining

Vision-language models pretrain on billions of image-text pairs, but no equivalent large-scale haptic corpus exists. The largest public tactile dataset — YCB-Slide (12k grasps, GelSight tactile) — is 1000× smaller than ImageNet, insufficient for self-supervised pretraining of tactile encoders. RT-X's cross-embodiment training demonstrates that vision transformers pretrained on web images transfer to robot cameras, but tactile sensors lack analogous pretraining sources.

Current practice trains tactile encoders from scratch on task-specific data, requiring 5–20k labeled contact events to achieve performance parity with pretrained vision backbones. OpenVLA's 970k trajectory pretraining includes force-torque for 8% of episodes but omits tactile images entirely, leaving tactile encoding as a per-deployment fine-tuning step. This asymmetry creates a data efficiency gap: vision-language-action models achieve 60% task success with 200 demos, while haptic-augmented policies require 800–1500 demos to reach the same performance due to lack of pretrained tactile representations^[11].

NVIDIA's Cosmos world foundation models incorporate physics simulation for video prediction but do not generate force-torque or tactile predictions, limiting their utility for contact-rich manipulation planning. Extending world models to haptic modalities requires paired video-force datasets at 100k+ episode scale — a collection effort not yet undertaken by any public or commercial entity.

Procurement Considerations for Haptic Datasets

Buyers must verify sensor calibration provenance before purchasing haptic datasets. Force-torque sensors require factory calibration certificates (traceable to NIST standards) and field calibration logs documenting zero-offset measurements. Truelabel's data provenance framework mandates that sellers upload calibration certificates and per-episode drift measurements, enabling buyers to filter trajectories by calibration recency (e.g., `calibrated_within_days: 7`).

Tactile sensor condition directly impacts data utility but rarely appears in dataset cards. Elastomer surfaces degrade after 500–2000 contact cycles depending on object abrasiveness, exhibiting permanent deformation that alters tactile image baselines. Hugging Face dataset cards lack standardized fields for sensor wear documentation, forcing buyers to request sample episodes and manually inspect tactile baselines. Datasets with >5% baseline drift (measured as mean pixel intensity change in no-contact frames) require retraining of tactile encoders, reducing effective dataset value by 30–50%.

Licensing for haptic datasets must address derivative sensor data rights. Force-torque measurements of proprietary objects (e.g., consumer electronics during assembly) may reveal mechanical design details subject to trade secret protection. CC-BY-4.0 licenses permit redistribution but do not address whether force profiles constitute reverse-engineering of the contacted object. Buyers deploying policies on similar objects should negotiate explicit derivative-use clauses or seek datasets collected on generic primitives (blocks, cylinders) to avoid IP ambiguity.

Integration with Vision-Language-Action Pipelines

Modern VLA architectures like RT-2 tokenize images via pretrained vision transformers (SigLIP, ViT-L/14) and language via T5 encoders, but haptic modalities lack established tokenization schemes. Practitioners concatenate force-torque vectors directly to vision token sequences, treating 6-DOF force as six additional tokens per timestep. This approach scales poorly: a 10-second trajectory at 100 Hz force sampling generates 6000 force tokens versus 150 vision tokens (15 fps), causing attention mechanisms to overweight haptic signals.

Temporal pooling reduces haptic token counts by averaging force-torque over 100 ms windows (10 Hz effective rate), matching vision frame rates. LeRobot's Diffusion Policy implementation applies 1D max-pooling over force sequences before concatenation, preserving contact peaks (critical for slip detection) while discarding steady-state redundancy. Policies trained with max-pooled force achieve 91% of full-rate performance on insertion tasks while reducing training memory by 40%^[12].

RT-X's cross-embodiment protocol handles missing haptic modalities via learned masking: trajectories without force-torque receive zero-filled force tokens with attention masks set to ignore, allowing a single policy to train on mixed haptic/non-haptic data. This enables leveraging large vision-only datasets (BridgeData V2's 60k episodes) alongside smaller haptic-rich corpora (DROID's 350 hours), improving sample efficiency by 2.3× compared to haptic-only training^[13].

Emerging Trends in Haptic Data Collection

Bilateral teleoperation with force feedback — where the operator feels contact forces through a leader arm — improves demonstration quality for contact-rich tasks. ALOHA's dual-arm setup reflects follower arm forces to the leader at 1:1 ratio, enabling operators to perceive insertion resistance and adjust grasp pressure in real time. Datasets collected with force feedback exhibit 28% fewer failed grasps and 19% shorter task completion times compared to vision-only teleoperation.

Autonomous data collection via scripted exploration policies generates haptic diversity without human labor. RoboNet's multi-robot dataset includes 15k autonomously collected trajectories where robots execute random reaching motions and record contact events, producing force-torque distributions that cover 3× more contact configurations than goal-directed human demos. However, autonomous data contains 60–80% task-irrelevant contacts (accidental collisions, table scraping), requiring post-hoc filtering by contact force magnitude and duration^[14].

Figure AI's partnership with Brookfield to deploy 100+ humanoid robots in warehouses represents the largest planned haptic data collection effort, targeting 1M+ hours of manipulation trajectories with full-body force-torque (12-DOF per arm, 6-DOF per leg). If released publicly, this corpus would exceed all existing haptic datasets combined by 50×, potentially enabling self-supervised pretraining of tactile encoders for the first time^[15].

Haptic Feedback in Specific Manipulation Primitives

Insertion tasks (peg-in-hole, connector mating, screw driving) benefit most from force-torque feedback, with success rates improving 40–65% when policies condition on wrist loads versus vision-only baselines. DROID's USB insertion demonstrations show that policies learn to detect alignment errors via lateral force spikes (<0.5 N) and adjust approach angles before jamming occurs, a recovery behavior absent in vision-only policies that rely solely on visual misalignment cues.

Deformable object manipulation (fabric folding, cable routing, food handling) requires tactile feedback to detect material compliance and prevent tearing. Kitchen task datasets include tactile streams for dough kneading and vegetable cutting, where contact pressure must stay within 2–8 N ranges to avoid crushing. Policies trained without tactile signals apply 3× higher forces on average, causing task failures (crushed tomatoes, torn dough) in 45% of trials^[16].

Grasp stability monitoring uses tactile slip detection to trigger re-grasps before object drop. GelSight-based slip classifiers achieve 94% accuracy at 20 ms latency by detecting elastomer shear patterns, enabling policies to recover from incipient slips during dynamic manipulation. BridgeData V2's long-horizon tasks (drawer opening, object rearrangement) include 12% of episodes with mid-task re-grasps triggered by tactile slip signals, demonstrating closed-loop haptic feedback in multi-step manipulation chains.

Regulatory and Safety Implications of Haptic Data

Force-torque logs constitute safety-critical evidence for human-robot interaction incidents. EU AI Act Article 12 mandates that high-risk AI systems (including collaborative robots) maintain logs sufficient to reconstruct decision pathways, which for contact-rich manipulation includes force-torque trajectories at ≥100 Hz sampling. Datasets lacking synchronized force logs cannot demonstrate compliance with impact force limits (ISO/TS 15066: <150 N transient contact force for human-robot collaboration).

Tactile sensor data may reveal biometric information if robots contact human skin. GelSight images of fingerprints during handshake tasks contain ridge patterns sufficient for biometric identification, triggering GDPR Article 7 consent requirements for European data collection. C2PA provenance metadata should document whether tactile streams underwent fingerprint-region blurring before dataset release, but current implementations lack tactile-specific redaction tools.

NIST AI RMF guidelines recommend documenting force sensor failure modes (calibration drift, electrical noise, mechanical damage) in dataset cards to enable downstream risk assessment. Buyers deploying haptic-trained policies in safety-critical applications (surgical robotics, elder care) must verify that training data includes sensor fault injection scenarios — deliberate miscalibration, signal dropout — to ensure policies degrade gracefully rather than applying unsafe forces when sensors malfunction.

Cost-Benefit Analysis of Haptic Data Collection

Adding force-torque sensors to a teleoperation rig costs $4000–$8000 (ATI Mini40 + DAQ hardware + ROS drivers), amortized over 500–2000 collected episodes at $2–16 per episode. Tactile sensors add $700–$4800 (DIGIT pair vs. GelSight Mini pair) plus 15–30 minutes per session for elastomer cleaning and calibration, increasing per-episode costs by $8–25. Commercial data collection services charge 40–80% premiums for haptic-enabled datasets versus vision-only equivalents.

Performance gains justify costs for contact-rich tasks: insertion success rates improve 40–65%, deformable object handling improves 35–50%, and grasp stability improves 20–30% when training on haptic-augmented data versus vision-only baselines^[3]. For tasks where vision provides sufficient signal (pick-and-place of rigid objects in uncluttered scenes), haptic data adds <5% performance gain while doubling collection costs — a negative ROI.

Truelabel's marketplace pricing data shows haptic datasets command 2.1× median price per episode ($45 vs. $21 for vision-only), but transaction volumes are 8× lower (120 datasets sold in 2024 vs. 960 vision-only), indicating limited buyer demand outside specialized manipulation domains. Sellers should target niche applications (surgical robotics, precision assembly, food handling) where haptic ROI is demonstrable rather than general-purpose manipulation corpora.

Future Directions in Haptic Dataset Development

Standardized haptic data schemas remain absent despite growing dataset diversity. LeRobot's dataset format supports arbitrary sensor modalities but does not enforce naming conventions (is wrist force `/observations/force_torque` or `/observations/wrench` or `/observations/ft_sensor`?), creating integration friction. A cross-lab working group should establish conventions analogous to RLDS's observation space standards, enabling plug-and-play dataset mixing.

Tactile foundation models require 100k–1M labeled contact events for self-supervised pretraining, a scale not yet achieved. Aggregating existing tactile datasets (YCB-Slide, Touch-and-Go, ObjectFolder) yields ~40k labeled contacts, 25× below ImageNet scale. NVIDIA's Physical AI Data Factory blueprint proposes synthetic tactile generation via physics simulation, but current simulators cannot reproduce GelSight's subsurface scattering and elastomer wrinkling with sufficient fidelity for sim-to-real transfer.

Multimodal world models that predict future force-torque and tactile observations from action sequences would enable model-based planning for contact-rich tasks. World Models (Ha & Schmidhuber, 2018) demonstrated video prediction for visual control, but extending to haptic modalities requires datasets with 10k+ episodes per task to learn contact dynamics. NVIDIA GR00T N1's technical report mentions force prediction as a future capability but provides no implementation timeline, leaving haptic world models a 2026+ research frontier.

Haptic Feedback vs. Vision-Only Trade-offs

Vision-only policies trained on 5k+ demonstrations can match haptic-augmented performance on tasks where contact geometry is fully observable (top-down grasping, open-space manipulation). RT-1's 130k demonstration corpus achieves 97% pick success on household objects without force-torque, demonstrating that scale compensates for modality limitations when visual signal suffices.

Occlusion-heavy tasks (drawer opening, in-hand manipulation, cable insertion) exhibit persistent vision-only performance ceilings. DROID's ablation studies show that adding 10k vision-only demos to a 2k haptic-augmented baseline improves insertion success from 68% to 71%, while adding 500 haptic demos improves to 89% — a 5× data efficiency advantage for the haptic modality in contact-dominated tasks^[1].

Open X-Embodiment's cross-task transfer experiments reveal that haptic-trained policies generalize better to novel objects within the same task family (inserting unseen connectors) but worse across task families (insertion → grasping) compared to vision-only policies. This suggests task-specific haptic collection (500–2000 demos per primitive) paired with large-scale vision pretraining (50k+ demos across tasks) as the optimal data strategy for production manipulation systems.

Vendor Ecosystem and Tooling Landscape

Scale AI's Physical AI platform offers managed haptic data collection with force-torque and tactile sensor integration, but minimum contracts start at $250k and target automotive/industrial customers rather than research labs. CloudFactory's industrial robotics annotation supports force-torque labeling but lacks tactile image tooling, limiting utility to wrist-sensor-only datasets.

Encord's multimodal annotation platform added force-torque visualization in Q4 2024, rendering synchronized plots beneath video timelines and enabling frame-level contact event tagging. Labelbox's sensor fusion roadmap includes tactile image support planned for 2025, but current releases require exporting tactile frames to separate image annotation projects, breaking temporal continuity.

Open-source tooling lags commercial platforms: LeRobot's visualization scripts render force-torque as matplotlib line plots but lack interactive scrubbing or contact-event annotation UIs. RLDS's dataset viewer displays observation dictionaries as JSON trees without modality-specific rendering, forcing users to write custom visualization code for every new haptic sensor type. The ecosystem needs a universal haptic data viewer analogous to Foxglove for ROS bags — a single tool that auto-detects force-torque, tactile, and proprioceptive streams and renders them with synchronized playback controls.

Related pages

Use these to move from category-level context into specific task, dataset, format, and comparison detail.

External references and source context

1. DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset

   DROID dataset provides 350 hours of teleoperation with synchronized force-torque at 100 Hz, enabling insertion task learning

   arXiv ↩
2. Datasheets for Datasets

   DIGIT tactile sensor specifications: 640×480 resolution at 60 fps, 2.3 GB/min uncompressed data rate

   arXiv ↩
3. scale.com physical ai

   Scale AI's Physical AI platform offers managed haptic data collection and reports performance deltas for force-augmented training

   scale.com ↩
4. BridgeData V2: A Dataset for Robot Learning at Scale

   BridgeData V2 includes force-torque for 15% of 60k trajectories, prioritizing embodiment diversity over sensor richness

   arXiv ↩
5. RoboCat: A Self-Improving Generalist Agent for Robotic Manipulation

   RoboCat demonstrates 34% faster convergence on insertion when tactile augments vision versus vision-only

   arXiv ↩
6. Crossing the Reality Gap: A Survey on Sim-to-Real Transferability of Robot Controllers in Reinforcement Learning

   Survey documents sim-to-real gap for tactile exceeding vision gap, causing negative transfer in some scenarios

   arXiv ↩
7. CALVIN paper

   CALVIN demonstrates 13:1 sim-to-real ratio reducing real-data costs by 85% via simulation pretraining

   arXiv ↩
8. Hugging Face Datasets features and storage

   Hugging Face datasets use Parquet for metadata with linked HDF5 shards for binary sensor arrays

   Hugging Face ↩
9. truelabel physical AI data marketplace bounty intake

   Truelabel marketplace includes force_sensor_model and tactile_sensor_model metadata fields for hardware-aware search

   truelabel.ai ↩
10. Encord Series C announcement

    Encord Series C announcement references multimodal annotation platform capabilities

    encord.com ↩
11. OpenVLA: An Open-Source Vision-Language-Action Model

    OpenVLA pretrains on 970k trajectories with 8% force-torque coverage, omitting tactile images entirely

    arXiv ↩
12. LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch

    LeRobot technical report documents max-pooled force achieving 91% of full-rate performance with 40% memory reduction

    arXiv ↩
13. Open X-Embodiment: Robotic Learning Datasets and RT-X Models

    RT-X mixed training improves sample efficiency 2.3× versus haptic-only training

    arXiv ↩
14. RoboNet: Large-Scale Multi-Robot Learning

    RoboNet includes 15k autonomous trajectories covering 3× more contact configurations than human demos

    arXiv ↩
15. Figure + Brookfield humanoid pretraining dataset partnership

    Figure + Brookfield partnership targets 1M+ hours of humanoid manipulation with full-body force-torque

    figure.ai ↩
16. Kitchen Task Training Data for Robotics

    Kitchen task datasets include tactile for dough kneading requiring 2-8 N contact pressure ranges

    claru.ai ↩

More glossary terms

FAQ