Teaching Robots to Say 'I Don't Know': SENTINEL for Uncertainty-Aware SLAM
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "Teaching Robots to Say 'I Don't Know'".
Dev: Here's the gist of the paper. Low-cost 2D LiDARs lack the intensity channels that more expensive sensors use to diagnose measurement failures, yet they are the default sensor on thousands of educational and budget robotics platforms.
Rosa: SENTINEL, a trainingfree, label-free reliability estimation framework that gives rangeonly LiDAR a diagnostic channel it does not natively possess.
Dev: The core problem addressed SLAM systems fail silently when sensor assumptions break, leading to confidently wrong pose estimates and phantom map structures with no error signal.
Rosa: A distinct category of failure is physical sensor failure where material properties cause the sensor to return structurally invalid data, such as transparent surfaces returning infinity or specular reflectors redirecting the beam.
Paper summary: Rosa: So wrapping up "Teaching Robots to Say 'I Don't Know': SENTINEL for Uncertainty-Aware SLAM," the authors are essentially showing how range-only LiDAR can be given a reliability metric without needing intensity channels.
Dev: They present this framework that fuses geometry based scan statistics with cross modal depth consistency to produce that score R between zero and one, and when R drops below a threshold, they suppress the corrupted scans and fall back to wheel odometry.
Taro: The key finding is the three point eight times separation between the clean condition minimum and the glass condition minimum in spatial reliability maps which proves it can detect physical failures <ref:2606.04853#pg3>.
Rosa: It’s about giving us a way to explicitly acknowledge perceptual uncertainty when that score R is below zero point three by suppressing corrupted scans and falling back to calibrated wheel odometry.
Rosa: The main implication is that for budget robotics platforms, adding a simple cross modal consistency check can provide a safety net against silent failures in SLAM.
Dev: It's about getting better at handling the uncertainty when the world misbehaves, not just running a faster loop rate or some other tuning adjustment.
Conclusion: Rosa: So we've been talking about SENTINEL, this framework that helps low cost range-only LiDAR systems know when they might be giving us bad data by using a cross modal check with a camera.
Dev: Yeah, it’s about that uncertainty awareness you get when the sensor assumptions break and the system starts giving us wrong pose estimates.
Rosa: The authors call it SENTINEL, and what’s really interesting is how they tackle the problem of those silent failures where a robot just keeps mapping a phantom structure without any error signal.
Dev: They show this reliability estimation framework that gives range-only LiDAR a diagnostic channel it doesn't have natively by fusing geometry stats with cross modal depth consistency.
Rosa: The big result they highlight is this separation, about three point eight times between the cleanest conditions and when you introduce things like glass surfaces into the map.
Dev: That separation proves it can actually detect those physical failures, like transparent surfaces returning infinity which would otherwise just confuse a standard SLAM system.
Rosa: It’s about explicitly acknowledging that uncertainty when that reliability score drops below zero point three, so the robot suppresses the bad scans and switches to wheel odometry.
Dev: That fallback is crucial for safety because it prevents silent map corruption and keeps the navigation stable while the system figures out what's actually happening.
Rosa: So, this means we’re moving towards systems that can be more honest about their own knowledge, even when using cheaper sensors on budget platforms.
Dev: It suggests a path forward where sensor fusion isn't just about combining data for accuracy but also for building a confidence score in what the robot is seeing.
Rosa: But there are still things to look at, like how they tuned those thresholds and what happens when you use modalities that fail together, which points us toward future work on learned thresholds and other ways to build this resilience.
Abhishek S, Badrikanath Praharaj, Sreeram M.V
BuildMachineLabs
cs.RO
Submitted: 2026-06-03
Updated: 2026-11-09
Comments: 6 pages, 10 figures, 3 tables, Accepted at the Uncertainty in Open-World Robotics Workshop, held in conjunction with the 2026 IEEE International Conference on Robotics and Automation (ICRA 2026)
Code: https://github.com/MichaelGrupp/evo
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 83/100
The gist: Low-cost 2D LiDARs lack the intensity channels that more expensive sensors use to diagnose measurement failures…
Key concepts
- Geometric Score (Rgeo)
- This score is calculated from raw scan statistics using beam validity ratio and range variance. It measures the quality of the LiDAR scan based on how well individual beams are performing. A low Rgeo indicates poor geometric quality, often due to obstructions or sensor issues.
- Cross-Modal Consistency (Rcross)
- This signal is derived by projecting LiDAR ranges into a depth image and counting agreements between the two modalities. It checks if the LiDAR's range measurements align with what the camera perceives as depth. High consistency suggests both sensors agree on the environment's structure.
- SENTINEL Framework
- SENTINEL is a training-free method that fuses geometric scores and cross-modal consistency to produce a per-scan reliability score R. This score helps SLAM systems suppress corrupted scans when uncertainty is high, preventing silent map corruption from physical sensor failures.
- Sensor Independence Finding
- The study found that effective cross-modal reliability requires modalities that fail independently. Because the depth camera and LiDAR share an optical blind spot for infrared materials, this suggests future fusion systems should consider sensors with different failure modes.
Terminology
Summary
Low-cost 2D LiDARs lack the intensity channels that more expensive sensors use to diagnose measurement failures, yet they are the default sensor on thousands of educational and budget robotics platforms. p. 0 *
BuildMachineLabs
mvsreeramblr@gmail. com
Abstract—Low-cost 2D LiDARs lack the intensity channels that
more expensive sensors use to diagnose measurement failures,
yet they are the default sensor on thousands of educational and
budget robotics platforms. com
Abstract—Low-cost 2D LiDARs lack the intensity channels that
more expensive sensors use to diagnose measurement failures,
yet they are the default sensor on thousands of educational and
budget robotics platforms. We present SENTINEL, a trainingfree, label-free reliability estimation framework that gives rangeonly LiDAR a diagnostic channel it does not natively possess.Where the paper says it
Where the paper says it
The core problem addressed
SLAM systems fail silently when sensor assumptions break, leading to confidently wrong pose estimates and phantom map structures with no error signal. A distinct category of failure is physical sensor failure where material properties cause the sensor to return structurally invalid data, such as transparent surfaces returning infinity or specular reflectors redirecting the beam. p. 0 A distinct, under-explored category is physical sensor
failure: the environment has adequate geometry, but material
properties cause the sensor to return structurally invalid data. Transparent surfaces transmit the 905 nm laser wavelength
used by low-cost LiDARs, returning infinity. These failures are invisible to degeneracy detectors
because scan matching is not ill-conditioned—the data is
simply wrong. Range-only LiDARs (RPLidar A1/A2, YDLidar X4/A1)
provide no intensity, no dual return, and no built-in diagnostics.Where the paper says it
Where the paper says it
How the training signal is built
SENTINEL fuses geometry-based scan statistics with cross-modal depth consistency between LiDAR and an RGB-D camera to produce a per-scan score R ∈ [0, 1]. This fusion involves two components:
-
Geometric Score (Rgeo): This is calculated from raw scan statistics using the formula Rgeo = α v + (1 − α) s. The geometric score is based on beam validity ratio and range variance. p. 0
Where the paper says it
We address this gap with 3 contributions: 1) SENTINEL: a training-free, label-free reliability framework running in real time embedded hardware. SENTINEL fuses a geometry-based score (Rgeo, 10 Hz) from raw scan statistics (beam validity ratio, range variance) with a crossmodal consistency signal (Rcross, 5 Hz) from an RGB-D camera, producing a fused score R at 11.
-
Cross-Modal Consistency (Rcross): This signal is derived by projecting the LiDAR range into the depth image and counting agreements between the two modalities.
Where the gains come from
The system's effectiveness stems from its ability to detect failures that are invisible in standard simulation, as physical failure modes like transparent surfaces are not reproducible in Gazebo’s default ray-casting model. The validation on real hardware across five surface conditions—glass, mirror, shining paper, and mixed—shows a significant separation between the clean-condition minimum and the glass-condition indicating SENTINEL correctly classifies affected tiles as reject or noise. p. 0 Spatial reliability maps
across five surface conditions like glass, mirror, shining paper and
mixed (mirror & shining paper) show a significant separation
between the clean-condition minimum and the glass-condition
indicating SENTINEL correctly classifies affected tiles as reject
or noise. Since none of the failure modes occur in simulation, we
validate the system on real hardware.Where the paper says it
Sensor independence finding
One key finding is that effective cross-modal reliability requires modalities that fail independently, constraining future sensor selection for fusion-based reliability systems. Specifically, two infrared sensors—Depth camera and LiDAR—share an optical blind spot for IR-transparent materials. p. 0 2) A sensor-independence finding: effective cross-modal
reliability requires modalities that fail independently. Two
infrared sensors—Depth camera and LiDAR—share an
optical blind spot for IR-transparent materials.Where the paper says it
Limitations
The authors note several limitations, including the fact that Rcross is unreliable for IR-transparent surfaces due to the shared infrared blind spot, requiring a non-infrared cross-modal source. Furthermore, all weights and thresholds were empirically tuned from representative runs rather than being formally derived. p. 5 2) All weights (α = 0. 5) and thresholds were empirically tuned from representative runs; formal sensitivity
analysis and automatic threshold learning are future work. 5) and thresholds were empirically tuned from representative runs; formal sensitivity
analysis and automatic threshold learning are future work. 3) Odometry fallback accumulates drift over extended failure
zones (> 2 m), and systematic evaluation of downstream
navigation performance remains future work.Where the paper says it
Where the paper says it
Conclusion
SENTINEL explicitly acknowledges perceptual uncertainty when R < 0.3 by suppressing corrupted scans and falling back to calibrated wheel odometry, which prevents silent map corruption. p. 0 SENTINEL fuses geometry-based scan statistics with cross-modal
depth consistency between LiDAR and an RGB-D camera to
produce a per-scan score R ∈ [0, 1]. When R drops below threshold, corrupted scans are suppressed and the robot falls back to
calibrated wheel odometry, preventing silent SLAM corruption. com
Abstract—Low-cost 2D LiDARs lack the intensity channels that
more expensive sensors use to diagnose measurement failures,
yet they are the default sensor on thousands of educational and
budget robotics platforms. We present SENTINEL, a trainingfree, label-free reliability estimation framework that gives rangeonly LiDAR a diagnostic channel it does not natively possess. We validate on GEFIER R1, a 4 wheel skid-steer robot equipped
with an RPLidar A2M12 and Intel RealSense D435i, navigating a
185 cm×245 cm arena with controlled transparent and reflective
failure elements on a central obstacle. Spatial reliability maps
across five surface conditions like glass, mirror, shining paper and
mixed (mirror & shining paper) show a significant separation
between the clean-condition minimum and the glass-condition
indicating SENTINEL correctly classifies affected tiles as reject
or noise. Two findings extend beyond the specific system: cross-modal
consistency requires sensors with truly independent failure
modes, and these failure modes are invisible in standard
Gazebo simulation. Future work will address learned thresholds, visual SLAM fallback, downstream navigation evaluation, and cross-platform validation.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
[1] K. Ebadi et al. p. 5 REFERENCES
[1] K. Ebadi et al. Ebadi et al. , “DARE-SLAM: Degeneracy-Aware and Resilient Loop
Closing in Perceptually-Degraded Environments,” J.Where the paper says it
Where the paper says it
[2] W. Wu, C. Chen, B. Yang et al. p. 5 , “DARE-SLAM: Degeneracy-Aware and Resilient Loop
Closing in Perceptually-Degraded Environments,” J. Yang et al. Yang et al. , “DALI-SLAM: Degeneracy-Aware
LiDAR-Inertial SLAM with Novel Distortion Correction and Accurate
Multi-Constraint Pose Graph Optimization,” ISPRS J. , “DALI-SLAM: Degeneracy-Aware
LiDAR-Inertial SLAM with Novel Distortion Correction and Accurate
Multi-Constraint Pose Graph Optimization,” ISPRS J. Photogramm. Photogramm. Remote Sens. Remote Sens. 92–110, 2025.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[3] K. Weerakoon et al. p. 5 92–110, 2025. Weerakoon et al. Weerakoon et al. , “TOPGN: Real-time Transparent Obstacle Detection using Lidar Point Cloud Intensity for Autonomous Robot Navigation,” arXiv:2408. , “TOPGN: Real-time Transparent Obstacle Detection using Lidar Point Cloud Intensity for Autonomous Robot Navigation,” arXiv:2408. 05608, 2024.Where the paper says it
Where the paper says it
Where the paper says it
[4] R. Zhao and A. p. 1 TOPGN [3]
uses per-beam intensity summed over multi-layer grids. Zhao
and Schwertfeger [4] classify reflective surfaces via dualreturn patterns. Zhao and A. Schwertfeger, “Reflection Detection via Plane Optimization,” in Proc. Schwertfeger, “Reflection Detection via Plane Optimization,” in Proc. IROS Workshop, 2024.Where the paper says it
Where the paper says it
Where the paper says it
[5] P. Foster et al. p. 1 Zhao
and Schwertfeger [4] classify reflective surfaces via dualreturn patterns. Foster et al. Foster et al. , “The Reflectance Field Map: Mapping Glass and
Specular Surfaces in Dynamic Environments,” in Proc. , “The Reflectance Field Map: Mapping Glass and
Specular Surfaces in Dynamic Environments,” in Proc. IEEE ICRA,
2023, pp. IEEE ICRA,
2023, pp. 4230–4236.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[6] T. Shan et al. p. 5 4230–4236. Shan et al. Shan et al. , “LVI-SAM: Tightly-Coupled Lidar-Visual-Inertial Odometry via Smoothing and Mapping,” in Proc. , “LVI-SAM: Tightly-Coupled Lidar-Visual-Inertial Odometry via Smoothing and Mapping,” in Proc. IEEE ICRA, 2021, pp. IEEE ICRA, 2021, pp. 5692–
5698.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[7] E. Chen et al. p. 5 5692–
5698. Chen et al. Chen et al. , “ALTER: Learning-on-the-Drive Self-supervised Adaptation of Visual Offroad Traversability Models,” in Proc. , “ALTER: Learning-on-the-Drive Self-supervised Adaptation of Visual Offroad Traversability Models,” in Proc. IEEE/RSJ IROS,
2024.Where the paper says it
Where the paper says it
Where the paper says it
[8] M. Labbé and F. p. 5 IEEE/RSJ IROS,
2024. Labbé and F. Labbé and F. Michaud, “RTAB-Map as an Open-Source Lidar
and Visual SLAM Library for Large-Scale and Long-Term Online
Operation,” J. Michaud, “RTAB-Map as an Open-Source Lidar
and Visual SLAM Library for Large-Scale and Long-Term Online
Operation,” J. Field Robot. Field Robot. 416–446, 2019.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[9] M. Grupp, “evo: Python Package for the Evaluation of Odometry and SLAM,” https://github. p. 5 416–446, 2019. Grupp, “evo: Python Package for the Evaluation of Odometry and
SLAM,” https://github. Grupp, “evo: Python Package for the Evaluation of Odometry and
SLAM,” https://github. com/MichaelGrupp/evo, 2017.Where the paper says it
Where the paper says it
[10] C. Cadena et al. p. 5 com/MichaelGrupp/evo, 2017. Cadena et al. Cadena et al. , “Past, Present, and Future of Simultaneous Localization
and Mapping: Toward the Robust-Perception Age,” IEEE Trans. , “Past, Present, and Future of Simultaneous Localization
and Mapping: Toward the Robust-Perception Age,” IEEE Trans. 1309–1332, 2016.Where the paper says it
Where the paper says it
Where the paper says it
[11] S. Thrun, W. Burgard, and D. p. 5 1309–1332, 2016. Burgard, and D. Burgard, and D. Fox, Probabilistic Robotics. Fox, Probabilistic Robotics. Cambridge,
MA: MIT Press, 2005.Where the paper says it
Where the paper says it
Where the paper says it
[12] S. Macenski and I. p. 5 Cambridge,
MA: MIT Press, 2005. Macenski and I. Macenski and I. Jambrecic, “SLAM Toolbox: SLAM for the Dynamic
World,” J. Jambrecic, “SLAM Toolbox: SLAM for the Dynamic
World,” J. Open Source Softw. Open Source Softw. 2783, 2021.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[13] S. Macenski et al. p. 5 Cambridge,
MA: MIT Press, 2005. Macenski and I. Cambridge,
MA: MIT Press, 2005. Macenski and I. Macenski et al. , “The Marathon 2: A Navigation System,” in Proc. , “The Marathon 2: A Navigation System,” in Proc. IEEE/RSJ IROS, 2020, pp. IEEE/RSJ IROS, 2020, pp. 2718–2725.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[14] T. Moore and D. p. 5 2718–2725. Moore and D. Moore and D. Stouch, “A Generalized Extended Kalman Filter
Implementation for the Robot Operating System,” in Proc. Stouch, “A Generalized Extended Kalman Filter
Implementation for the Robot Operating System,” in Proc. (IAS), 2014.Where the paper says it
Where the paper says it
Where the paper says it
[15] S. Macenski et al., “Robot Operating System 2: Design, Architecture, and Uses in the Wild,” Sci. p. 5 Macenski et al. , “Robot Operating System 2: Design, Architecture,
and Uses in the Wild,” Sci.Where the paper says it
[16] AutoMind Dynamics, “Products and Technical Resources,” including product specifications, brochures, and demonstration materials. p. 5 , “Robot Operating System 2: Design, Architecture,
and Uses in the Wild,” Sci. [16] AutoMind Dynamics, “Products and Technical Resources,” including
product specifications, brochures, and demonstration materials.Where the paper says it
Table I
| Mode | Gazebo | Real hardware |
|---|---|---|
Glass Ray hits surface; valid distance Laser transmits through; returns ∞
| Metallic Ray hits surface; valid distance | Laser scatters; false distance | Mirror Ray hits surface; false distance |
| USB dropout | Never occurs Occurs under bus contention infrastructure failures have no simulation analogue.
Table II
| Component Key | Specification Cost |
|---|---|
RPLidar A2M12 2D, 360°, 10 Hz, range-only 230
RealSense D435i RGB-D, 848×480, IR light 300
| Adafruit BNO055 9-DOF absolute orientation, 68 Hz | 35 |
| ESP32 MCU Dual-core 240 MHz, micro-ROS | 10 |
Total sensing stack cost is estimated at a total of five hundred seventy-five dollars.
Table III
| Condition | Obstacle material |
|---|---|
| 1 Clean baseline | Opaque white cardboard walls |
2 Glass (transparent) Glass 0.5 mm laser transmits through, returning infinity. p. 1 Gazebo’s raycaster treats every surface as an opaque diffuse reflector
(Table I). For glass, the laser transmits through in hardware,
returning infinity; Gazebo returns a valid distance.Where the paper says it
3 Mirror (specular) Glass mirror redirects the laser at the angle of incidence. p. 2 • Shining paper (diffuse reflective): metallic foil producing
scattered false returns. • Glass mirror (specular reflective): redirects the laser at the
angle of incidence.Where the paper says it
4 Shining (diffuse) Metallic paper produces scattered false returns. p. 2 5 mm): laser transmits through, returning ∞. • Shining paper (diffuse reflective): metallic foil producing
scattered false returns.Where the paper says it
| 5 Mixed | Mirror on two faces, shining paper on the remaining two.
Table IV
| Condition | Rgeo | Rcross | R |
|---|---|---|---|
Normal (no failure) 0.82 1.00 0.91.
Acrylic (all sides) 0. p. 3 Condition Rgeo Rcross R
Normal (no failure) 0. 91
Acrylic (all sides) 0. We address this gap with 3 contributions:
1) SENTINEL: a training-free, label-free reliability framework running in real time embedded hardware. SENTINEL
fuses a geometry-based score (Rgeo, 10 Hz) from raw scan
statistics (beam validity ratio, range variance) with a crossmodal consistency signal (Rcross, 5 Hz) from an RGB-D
camera, producing a fused score R at 11.Where the paper says it
Where the paper says it
Recovery (removed) 0. p. 3 91
Acrylic (all sides) 0. 45
Recovery (removed) 0.Where the paper says it
Table V
| Tile | Clean Rgeo | Glass Rgeo | Mirror Rgeo | Shining Rgeo | Mixed Rgeo Min/Mean|
|---|---|---|---|---|---|
T1–T3, T9–T10 > 0.89 across all conditions. p. 0 We validate on GEFIER R1, a 4 wheel skid-steer robot equipped
with an RPLidar A2M12 and Intel RealSense D435i, navigating a
185 cm×245 cm arena with controlled transparent and reflective
failure elements on a central obstacle. Spatial reliability maps
across five surface conditions like glass, mirror, shining paper and
mixed (mirror & shining paper) show a significant separation
between the clean-condition minimum and the glass-condition
indicating SENTINEL correctly classifies affected tiles as reject
or noise.Where the paper says it
T4–T8 Glass drops to Rgeo = 0. p. 4 7: Rgeo time-series, T1–T10: Glass. Rgeo drops from 0.Where the paper says it
| Failure zone mean | Clean min Rgeo = 0.91 at T8.
Glass failure zone Sustains Rgeo ∈ [0. p. 5 Spatial reliability maps across
five surface conditions demonstrate a 3. 8× Rgeo separation
between the clean-condition minimum and the glass-condition
minimum, with consistent detection across the failure zone.Where the paper says it
Mirror/Shining Produce intermediate degradation (Rgeo = 0. p. 3 This
matches the expected physics: the laser transmits through
glass, producing beam loss that directly reduces the validity
ratio v (Equation (1)). Mirror and shining paper produce intermediate degradation
(Rgeo = 0.Where the paper says it
Table VI
| Method | Failure type | Int. GPU Train | L-free S2R |
|---|---|---|---|
| DARE-SLAM [1] | Geom. deg. No No No | Yes 3k+ No No.
| TOPGN [3] | Transparent Yes No No | 3k+ Yes Yes.
| ALTER [7] | Sensor fail. No Yes Yes | 3k+ Yes No.
| LVI-SAM [6] | General deg. No No No | 3k+ No No.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Reliability estimation framework for range-only LiDARs can be implemented to give them a diagnostic channel they do not natively possess by fusing geometry-based scan statistics with cross-modal depth consistency between LiDAR and an RGB-D camera to produce a per-scan score R in [0, 1].
-
The improved system can suppress corrupted scans when R drops below a threshold, causing the robot to fall back to calibrated wheel odometry and preventing silent SLAM corruption.
-
The system can perform sensor independence findings by constraining future sensor selection for fusion-based reliability systems by verifying that modalities fail independently, such as ensuring two infrared sensors share an optical blind spot for IR-transparent materials.
Abstract
Low-cost 2D LiDARs lack the intensity channel that higher-end sensors use to diagnose measurement failures, yet they are widely used on educational and budget robotics platforms. We present SENTINEL, a training - free, label - free reliability estimation framework that gives range - only LiDAR an effective diagnostic signal. SENTINEL combines geometry-based scan statistics with cross - modal depth consistency between LiDAR and an RGB - D camera to compute a per - scan reliability score between 0 and 1. When the score falls below a threshold, corrupted scans are rejected and the robot falls back to calibrated wheel odometry, preventing silent SLAM corruption. We evaluate SENTINEL on a GEFIER R1 four - wheel skid-steer robot equipped with an RPLidar A2M12 and an Intel RealSense D435i in a 185 cm by 245 cm arena containing controlled transparent and reflective failure elements on a central obstacle. Spatial reliability maps across five surface conditions, including glass, mirror, shiny paper, and a mixed mirror and shiny-paper condition, show clear separation between clean and failure cases, allowing affected regions to be identified as reject or noise. Because these failure modes are absent in simulation, validation is performed entirely on real hardware.
Sources
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation