Audio-Text Cross-Attention Model Detects Ambivalence and Hesitancy
A new multimodal system jointly encodes speech and transcript to recognise subtle behavioural states of ambivalence and hesitancy without requiring aligned audio and text.
Research Digest
Preprints and papers, rewritten in plain English.
AD-Mamba Distills CNN and Transformer Strengths into Efficient Audio Model
A new distillation framework lets a compact Mamba state-space model match or beat larger Transformer audio classifiers at lower computational cost.
Targeted data collection and offline RL boost contact-rich assembly
A modular robot-learning pipeline that collects data only where contact matters achieves 96% success on long-horizon assembly tasks with far less manual effort.
ShlokaDhara Corpus Brings Meter-Aware Prosody to Sanskrit Text-to-Recitation
A new 43-hour Sanskrit speech corpus targets metrical correctness in verse synthesis, with XTTS-v2 scoring highest in early benchmarks.
WorldScape Policy 2.0 adds memory and event data to robot world models
A new world action model uses reasoning-augmented memory and a large event-grounded dataset to improve long-horizon robot planning and instruction following.
ZeroDex: Zero-Shot Dexterous Manipulation via 3D-Grounded VLMs
A new framework uses multi-view RGB images and vision-language model reasoning to plan long-horizon robot manipulation without demonstrations or task-specific training.
Test-Time Trajectory Optimization Improves Autonomous Driving Planners
A new method treats a planner's scorer as a reward function and searches for better trajectories at test time, improving performance across multiple driving benchmarks.
Neural operator learns noise-aware planners that beat MC baselines
A new neural operator, PUN, learns noise-aware value fields for planning in unseen geometries, outperforming baselines in 4D replanning while running faster.
AgenticRoboRL: post-training lifts robot planner success to 71.0%
A new framework applies coding-agent post-training to robot planners, lifting a 9B model's LIBERO-PRO success rate from 18.1% to 71.0% without external memory.
Barrier functions keep multirotor control authority within certified limits
A new allocation scheme uses control barrier functions to guarantee that overactuated multirotors retain enough wrench-rate authority, avoiding the failures of greedy and effort-minimizing approaches.
Predicting silent failures in learned inference under restricted access
A new score predicts when learned inference systems report intervals that miss the truth, but finds measurement access—not the learned prior—drives the risk.
Latent action priors improve torque-based legged robot learning
Injecting expert-demonstrated action patterns into deep reinforcement learning helps legged robots learn more natural and robust torque control with less data.
Mondrian-based prior helps robots predict unseen indoor space
A training-free method that generates map hypotheses to guide indoor robot exploration matches learned approaches on benchmark floor plans.
PhysCanvas world model aligns visual and tactile feedback for robot training
A vision-tactile world model that keeps simulated contact dynamics consistent with real outcomes improves robot policy evaluation and post-training.
Gibson Env: A Virtual World Built from Real Spaces for Embodied AI
Gibson Env addresses the sim-to-real gap by building a simulation environment from virtualized real-world spaces, with a mechanism to transfer trained models directly to physical settings.
Standards-Guided Generation for Code Documentation
A new approach ties code documentation generation to explicit quality standards and dependency structure rather than treating it as unconstrained text.
A Regret Bound for Particle-Filter Deep RL in Partially Observable Worlds
New work proves an end-to-end regret bound for a practical deep RL algorithm that uses particle filters for belief inference, making the trade-offs among sample size, particle budget, and planning depth explicit.
Safe Reinforcement Learning for Autonomous Roundabout Navigation
A new arXiv paper proposes a prescribed-performance-guided safe reinforcement learning approach to help autonomous vehicles handle the challenges of roundabout navigation.