LESS (Local Encoder for Spatial Sensing)

ee_pose · **LESS** ("More with LESS: Local Scene Representations for Tactile Imaging", RSS 2026) tackles artificial palpation — reconstructing what is *inside* a soft object purely by touching it. The motivating application is breast cancer screening, where over **40%** of cases are first detected by palpation. - **>800 hours** of robotic tactile interaction, released as **~66 GB** across three Zenodo records (poke 40.3 GB, primitive 19.3 GB, out-of-distribution bundle 6.4 GB). - Collected with a motorised lift plus a **Franka Panda** arm carrying a **gel-based tactile sensor**, executing systematic poke and sweep primitives over a library of modular silicone breast phantoms, each containing a harder spherical inclusion at varying shape and location. - Every phantom configuration is scanned in an **MRI machine**, so each tactile sequence is paired with a volumetric ground-truth label — a rare combination in robot tactile data. - Four splits support explicit generalisation tests: *poke-primitive* (training, single inclusions), *poke-big* (zero-shot to **4× larger** phantoms), *poke-multi* (zero-shot to **2–3 simultaneous inclusions**, never seen in training), and *handheld* (sensor held by a human, pose tracked by fiducials rather than robot kinematics). - The accompanying method is a grid of **GRU particle encoders with local receptive fields**, each reconstructing a local patch — compositional by construction, which is why single-inclusion training transfers zero-shot to multi-inclusion phantoms. - Supports 2D and 3D reconstruction plus spatial uncertainty estimation, and enabled the first **real-time hand-held** tactile imaging device with 3D output. Data is **CC BY 4.0**; the code is MIT.

**LESS** ("More with LESS: Local Scene Representations for Tactile Imaging", RSS 2026) tackles artificial palpation — reconstructing what is *inside* a soft object purely by touching it. The motivating application is breast cancer screening, where over **40%** of cases are first detected by palpation. - **>800 hours** of robotic tactile interaction, released as **~66 GB** across three Zenodo records (poke 40.3 GB, primitive 19.3 GB, out-of-distribution bundle 6.4 GB). - Collected with a motorised lift plus a **Franka Panda** arm carrying a **gel-based tactile sensor**, executing systematic poke and sweep primitives over a library of modular silicone breast phantoms, each containing a harder spherical inclusion at varying shape and location. - Every phantom configuration is scanned in an **MRI machine**, so each tactile sequence is paired with a volumetric ground-truth label — a rare combination in robot tactile data. - Four splits support explicit generalisation tests: *poke-primitive* (training, single inclusions), *poke-big* (zero-shot to **4× larger** phantoms), *poke-multi* (zero-shot to **2–3 simultaneous inclusions**, never seen in training), and *handheld* (sensor held by a human, pose tracked by fiducials rather than robot kinematics). - The accompanying method is a grid of **GRU particle encoders with local receptive fields**, each reconstructing a local patch — compositional by construction, which is why single-inclusion training transfers zero-shot to multi-inclusion phantoms. - Supports 2D and 3D reconstruction plus spatial uncertainty estimation, and enabled the first **real-time hand-held** tactile imaging device with 3D output. Data is **CC BY 4.0**; the code is MIT.

No decision-grade evidence captured yet.

**LESS** ("More with LESS: Local Scene Representations for Tactile Imaging", RSS 2026) tackles artificial palpation — reconstructing what is *inside* a soft object purely by touching it. The motivating application is breast cancer screening, where over **40%** of cases are first detected by palpation. - **>800 hours** of robotic tactile interaction, released as **~66 GB** across three Zenodo records (poke 40.3 GB, primitive 19.3 GB, out-of-distribution bundle 6.4 GB). - Collected with a motorised lift plus a **Franka Panda** arm carrying a **gel-based tactile sensor**, executing systematic poke and sweep primitives over a library of modular silicone breast phantoms, each containing a harder spherical inclusion at varying shape and location. - Every phantom configuration is scanned in an **MRI machine**, so each tactile sequence is paired with a volumetric ground-truth label — a rare combination in robot tactile data. - Four splits support explicit generalisation tests: *poke-primitive* (training, single inclusions), *poke-big* (zero-shot to **4× larger** phantoms), *poke-multi* (zero-shot to **2–3 simultaneous inclusions**, never seen in training), and *handheld* (sensor held by a human, pose tracked by fiducials rather than robot kinematics). - The accompanying method is a grid of **GRU particle encoders with local receptive fields**, each reconstructing a local patch — compositional by construction, which is why single-inclusion training transfers zero-shot to multi-inclusion phantoms. - Supports 2D and 3D reconstruction plus spatial uncertainty estimation, and enabled the first **real-time hand-held** tactile imaging device with 3D output. Data is **CC BY 4.0**; the code is MIT.

Open · CC-BY-4.0 · custom