CrystalAR: Autoregressive Generation with Space-Group Symmetry
A new deep generative model explicitly incorporates space-group symmetry into autoregressive crystal generation.
Research Digest
Preprints and papers, rewritten in plain English.
Client-side embedding transform aims to keep prompts private from hosted LLMs
A new arXiv paper explores sending transformed vectors instead of raw text to hosted language models, so prompts never leave the client.
Latent projection prunes noise to improve world-model planning
A post-hoc correction that restricts planning costs to low-uncertainty latent directions sharply improves navigation success in latent world models.
Closed-Path Analysis Improves Platoon Control at Unsignalized Intersections
A new distributed controller uses a closed-path model to prove that connected vehicle platoons can cross unsignalized intersections safely and smoothly.
Tactile Feedback Lets GPT-6-Astra Match Specialized Robot Policies
A general-purpose multimodal model, GPT-6-Astra, achieves 93.8% mean success on contact-rich manipulation tasks without robotics-specific pretraining or per-task fine-tuning.
ForeTac uses predicted touch to guide robot manipulation before contact
A new framework predicts tactile outcomes of candidate actions and uses that foresight to guide generative robot policies, sharply improving contact-rich manipulation success.
Distribution over plans improves vision-language-action robot control
A new method lets robot policies keep multiple valid plans in mind, improving steering and adaptation without retraining.
Fast Retargeting Pipeline Speeds Humanoid Motion Transfer
A new framework retargets human motion capture data to humanoid robots up to 136 times faster than prior pipelines while validating dynamic feasibility.
TaskWAM steers world-action models away from stage confusion
New stage-aware framework helps robotic world-action models generate futures that match the current task and progress, not just any plausible next step.
SPAR ranks push actions to speed robot navigation around movable obstacles
A search-trained ranker lets a pushing robot distinguish setup pushes from immediate openings, cutting planning time by 3.8x in simulated keyhole problems.
Benchmark Reveals Gap in LLM Long-Horizon Safety Planning
A new benchmark shows that large language models can often recognize unsafe actions but still struggle to plan safe multi-step tasks in hazardous household environments.
Thermal Control Strategy for Micro Lunar Rover HAKUTO
A study details how the HAKUTO rover's thermal design was modeled and tested to keep a sub-4kg lunar rover within safe temperatures.
GPT-6 Astra Shows Gaps Between Reasoning and Robot Control
A new arXiv study systematically benchmarks GPT-6 Astra's embodied capabilities and finds a clear divide between semantic understanding and physical execution in robotic manipulation.
Cosmos-Augmented Training Lifts Robot Navigation Success to 91.7%
A benchmark study shows NVIDIA's Cosmos sim-to-real augmentation nearly closes the gap between learned and production navigation on industrial mobile robots.
Driving VLAs see hazards but don't brake: a perception-action gap
A new study identifies a binding gap in vision-language-action driving models, where hazard perception fails to influence trajectory generation, and demonstrates a weight-free fix.
LLM Planning Hardness: Structural Complexity vs. Symbol Interpretability
A new arXiv study separates structural difficulty from semantic interpretability in LLM planning benchmarks.
Curved stripe markers improve monocular roll estimates for cylinders
By exploiting known surface curvature, a dual-colour stripe marker lets a single camera estimate roll on thin cylinders far more precisely than a flat marker.
DirectWAM: Single-Step Robot Control Without Diffusion Sampling
A new arXiv preprint shows that robot-control futures can be distilled into one-step predictors, cutting inference latency 28-fold while improving success rates.