Is LingBot-Touch state of the art?Pending benchmarks.Announced Nov 9, 2026; no independent results yet.
Robbyant's open JEPA-based tactile foundation model for dexterous manipulation. A shared Haptic Transformer learns unified representations of magnetic, piezoresistive and visuotactile signals from nearly 5,000 hours of real tactile recordings, then transfers to dexterous hands without extra representation training. Tiny and Small encoder checkpoints are public on Hugging Face and ModelScope under Apache 2.0, with inference code on GitHub.
The details so far
- Lab
- Robbyant
- Announced
- Nov 9, 2026
- API
- Not yet
- Open weights
- Published · Public on Hugging Face (robbyant/lingbot-touch-tiny, robbyant/lingbot-touch-small) and ModelScope; ungated download verified
- Parameters
- HaT-Tiny encoder about 5.46M parameters (Robbyant project page); Small uses a larger shared Transformer (embed dim 384 vs 192) — Robbyant does not state a Small parameter count on the project page
- Type
- JEPA-style tactile representation model: modality-specific tokenizers plus a shared Haptic Transformer with latent prediction from spatiotemporal tactile context
- Modalities
- Magnetic (3-axis), piezoresistive (scalar in Z), and visuotactile (scalar deformation/intensity in Z; raw RGB is not an input)
- Pretraining data
- 4,983 hours / 464,431 episodes / about 1.33B frames across piezoresistive (2,197 h), magnetic (1,136 h) and visuotactile (1,650 h) recordings from tactile gloves and robot hands (project page)
- Variants
- LingBot-Touch-Tiny and LingBot-Touch-Small frozen target encoders (config.json + encoder.pt each); inference code in GitHub Robbyant/lingbot-touch
- Transfer
- Robbyant says the encoder transfers to unseen tactile sensors and real dexterous hands (Sharpa Wave, XHand) without additional representation training
- Downstream use
- Token-level latent features for tactile perception and robot policies; Robbyant recommends attention-based querying over mean/max pooling
- License
- Apache License 2.0 (GitHub LICENSE and Hugging Face model cards)
- Code
- https://github.com/Robbyant/lingbot-touch (Apache-2.0)
- Related
- Separate from other LingBot family models (Vision, Depth, Map, VLA, VA, World, Video); those are not this page
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: pretrained on nearly 5,000 hours (4,983 h) of real tactile data spanning three sensor types (Robbyant project page / X).
Lab claim: evaluated across 26 real-robot tasks on Sharpa Wave and XHand with four action models (LingBot-VLA, π0.5, Diffusion Policy, T-Rex); adding frozen LingBot-Touch features during post-training raises mean success versus original and raw-tactile baselines (project page charts).
Lab claim: even the Tiny backbone (about 5.46M parameters) supports high-frequency downstream inference; scaling pretraining data toward 5K hours continues to improve tactile and real-robot results (project page).
Sources
LABLingBot-Touch project page (Robbyant)technology.robbyant.com/lingbot-touch · Nov 9, 2026LABRobbyant on X: LingBot-Touchx.com/robbyant_brain/status/2119870554056651004 · Nov 9, 2026REPORobbyant/lingbot-touch on GitHubgithub.com/Robbyant/lingbot-touch · Nov 9, 2026REPOlingbot-touch-small on Hugging Facehuggingface.co/robbyant/lingbot-touch-small · Nov 9, 2026REPOlingbot-touch-tiny on Hugging Facehuggingface.co/robbyant/lingbot-touch-tiny · Nov 9, 2026Tracking continues until an independent board lists it.