Calvin
PulseAugur coverage of Calvin — every cluster mentioning Calvin across labs, papers, and developer communities, ranked by signal.
6 day(s) with sentiment data
-
New TOPS method enhances multi-task robot manipulation with adaptive priors
Researchers have developed TOPS (Task-Oriented Priors and Schedules), a novel method for language-conditioned multi-task robot manipulation. TOPS adapts the flow source prior and interpolation schedule to individual tas…
-
Robotic manipulation research finds smaller models outperform large ones for control targets
A new research paper challenges the prevailing assumption that high-quality image editing leads to better robotic manipulation control targets. The study found that while large proprietary models like Gemini 3 Pro Image…
-
New PLaCA Framework Enhances Robot Action Generation by Predicting Future Latent Change
Researchers have developed PLaCA, a new framework for Vision-Language-Action (VLA) models designed to improve robot action generation by predicting future environmental changes at a patch level. This approach directly m…
-
New SpecWM framework uses spectral alignment for improved world modeling
Researchers have introduced SpecWM, a new framework for world modeling that uses spectral representation learning to guide what information a model preserves. By specifying desired relationships between observations thr…
-
GreenWAM: A Unified Foundation Model for Robotics Control and Prediction
Researchers have introduced GreenWAM, a novel world foundation model designed for robotics. This single model functions as both an action-conditioned world model and a policy, capable of operating across 20 diverse robo…
-
ForeAct3D framework grounds robot world models in planned actions
Researchers have developed ForeAct3D, a new framework designed to improve the world modeling capabilities of Vision-Language-Action (VLA) policies for robots. Unlike previous methods that forecast future observations in…
-
New D2R method combats robot policy forgetting using world models
Researchers have developed a new method called Dream to Remember (D2R) to address catastrophic forgetting in generalist robot policies. D2R continually steers a frozen policy towards new tasks while using a pretrained w…
-
New RAFC method improves robot manipulation with reliable future conditioning
Researchers have developed a new method called Reliability-Aware Future Conditioning (RAFC) to improve robot manipulation by making generated future task videos more useful. Traditional methods can be hindered by tempor…
-
New research aims to boost VLA model generalization and efficiency
Researchers are exploring new methods to improve the generalization capabilities of Vision-Language-Action (VLA) models in robotics. One approach, Equivariant Counterfactual Training (ECT), addresses instruction-action …
-
AtomicVLA framework enhances robotic skill learning with atomic abstraction
Researchers have introduced AtomicVLA, a novel framework designed to enhance the capabilities of Visual-Language-Action (VLA) models in robotics. This system addresses the limitations of current monolithic VLA models by…
-
New frameworks advance AI world models for simulation and robotics · 8 sources tracked
Several research papers introduce new frameworks and models for advancing world-action models, which are crucial for enabling AI agents to understand and interact with the physical world. World in World offers a trainin…
-
StreamPI framework gives VLA models temporal understanding of physical world
Researchers from Da Xiao Robotics and the University of Hong Kong have introduced StreamPI, a novel framework designed to imbue Vision-Language-Action (VLA) models with a temporal understanding of the physical world. Un…
-
DFM-VLA introduces iterative action refinement for robot manipulation
Researchers have introduced DFM-VLA, a novel approach for robot manipulation that utilizes discrete flow matching to iteratively refine action tokens. Unlike previous methods that fix tokens once generated, DFM-VLA mode…
-
New framework uses exocentric data to improve egocentric 3D hand pose forecasting
Researchers have developed Exo2EgoPose, a novel framework designed to improve the forecasting of egocentric 3D hand poses. This method leverages exocentric demonstrations to guide and compensate for the limited and dyna…
-
New Anchor-Align method boosts VLA policy generalization
Researchers have introduced Anchor-Align, a novel method to improve vision-language-action (VLA) policies by addressing issues with standard behavior cloning (BC) finetuning. BC finetuning can degrade the generalizabili…
-
New methods enhance VLM to VLA adaptation for robotics control · 2 sources tracked
Two new research papers propose methods to improve the adaptation of vision-language models (VLMs) into vision-language-action (VLA) models for robotics. The first paper introduces CLAP (Causal Language-Action Predictio…
-
ThinkProprio integrates robot state to improve VLA model attention and speed
Researchers have developed a novel approach called ThinkProprio for vision-language-action (VLA) models, which integrates proprioceptive data more effectively into the decision-making process. Unlike traditional methods…
-
New benchmarks and frameworks advance video world modeling
Researchers have introduced "ImageTime," a new benchmark designed to evaluate how well image generation models can understand and represent temporal changes. This benchmark assesses spatiotemporal consistency by requiri…
-
New MoLA method bridges robot video imagination and action execution
Researchers have developed a new method called MoLA (Mixture of Latent Actions) to improve robot manipulation by better utilizing predicted future video frames. MoLA transforms these imagined futures into executable act…