#1 ·
Learning Heterogeneous Tactile Representation for General Dexterous Manipulation
Bing Han, Yichuan Peng, Haoyu Dong, Yuanmeng Hu
Embodied AITactile SensingRobot Manipulation
Why it matters
Commercial tactile sensors are fragmented, each with its own strengths and blind spots, which locks datasets to specific hardware. A foundation model that reads them all in a common space means dexterous manipulation stacks can mix and match sensors as hardware matures instead of betting on one device. For anyone building robot hands, it is the clearest signal yet that cross-sensor data, not better hardware alone, is how tactile understanding scales.
LingBot-Touch is a tactile foundation model that learns a unified representation across piezoresistive, magnetic, and visuotactile sensors, decoupling touch understanding from any single sensor's physics and remaining extensible to new types. The resulting tactile tokens are plug-and-play: integrated into diffusion policies or vision-language-action models such as π0.5, they consistently improve manipulation dexterity without architecture changes. Simulations and real-robot deployments across multiple hardware platforms show the representation substantially improves complex contact-rich manipulation while staying computationally efficient.
arXiv | HuggingFace
#2 ·
WorldFoundry: Towards Unified Infrastructure for World Intelligence
Juanxi Tian, Jiadong Hou, Baiang Li, Xin Lin
World ModelsOpen SourceEvaluation Infrastructure
Why it matters
World models are currently judged by how good their outputs look, not by whether they actually simulate: persistent state and causally consistent responses to interventions are the hard part, and there is no shared way to test them. WorldFoundry makes that definition concrete and auditable, with every score traceable to recorded execution evidence. If you are building agents or simulators, it is the measurement layer the field has been missing, and a template for how evaluation claims should be separated from execution claims.
WorldFoundry is open-source infrastructure for world modeling that starts from an explicit definition: a generative world model is a predictive dynamical system with persistent, history-conditioned state that updates under actions and generates temporally coherent trajectories under observed and counterfactual interventions. It unifies model and benchmark discovery, runtime and asset preparation, execution orchestration, artifact management, and interactive visualization across video, 3D/4D, embodied action, and interactive systems, with modality-specific adapters preserving native semantics. The catalog contains 270 model entries, 74 benchmarks, and 62 registered metric entries, and explicit manifests with evidence-aware scorecards keep execution readiness separate from evaluative claims.
arXiv | HuggingFace
#3 ·
Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery
Weida Wang, Benteng Chen, Yigeng Jiang, Shuchen Pu
AI for ScienceAgentic SystemsMetacognition
Why it matters
Most AI-scientist systems generate hypotheses and stop; the open question is whether an agent can govern its own research process. Sage-Taiyi's metacognitive loop — monitoring, revising, and steering its own evidence and assumptions — is the first architecture where that pattern clearly beats the strongest baseline at scale. And the wet-lab results show its proposals survive contact with reality, which is what separates a leaderboard score from a usable research copilot.
Sage-Taiyi is a metacognition-inspired architecture in which the system explicitly monitors and controls its own scientific process, revising its understanding of evidence, assumptions, and open questions, then using those assessments to adapt its research targets and select reusable reasoning methods. On Taiyi-Bench, a co-scientist ideation benchmark of 265 tasks from 50 papers across ten disciplines, it ranks first on all eight historical rediscovery and open-ended ideation metrics, scoring 54.0 overall versus 47.9 for the strongest baseline, AI Scientist-v2, and also performs well on two external ideation benchmarks. Five expert-guided collaborations connected its proposals to real validation: cobalt-catalyzed carbonylation yields rose from trace amounts to 84 percent, and photocatalytic allylic amination from 13 to 69 percent.
arXiv | HuggingFace
Outlook
The common thread is the field moving from impressive outputs to verifiable capability: touch that transfers across hardware, world models that can be audited against interventions, and scientific agents that manage their own uncertainty. Expect follow-ups on cross-sensor foundation models for robot hands and on agentic systems closing the loop with real experimental validation.