Researchers developed ForceTwin, a system for creating physics-informed digital twins of articulated objects using instrumented human interaction. This allows for more accurate modeling of object dynamics, including inertia, friction, and state-dependent mechanism forces. The system improves upon existing digital-twin pipelines, achieving higher goal completion rates in impedance control tasks and enabling whole-body door-traversal policies in the real world.

Researchers propose a reinforcement learning framework, PA-RL, that uses artificial potential fields as an action representation for contact-rich robotic manipulation tasks. This approach allows the policy to adapt task strategy and low-level motion generation, reducing learning burden and improving performance. PA-RL achieves 100% evaluation success rate in simulation and demonstrates real-robot deployment feasibility.

KnowDemo is a framework for generating robot demonstrations from human videos by leveraging structured manipulation knowledge. It uses a vision-language model to extract task conditions, demonstration references, and permissible execution variations, then translates this knowledge into executable demonstrations through motion planning and simulation. This approach improves behavioral diversity and demonstration generation efficiency.

Researchers introduce Latent Evolution Operator Network (LEON), a novel architecture for modeling latent evolution in World Action Models (WAMs) that improves closed-loop performance and robustness in robot policies. This development is relevant to people building and operating AI agents, particularly those using WAMs, as it offers a new approach to transition realization.

Researchers developed SimHum, a co-training recipe that leverages simulation and human demonstrations to improve data efficiency and scene generalizability in bimanual manipulation tasks. This approach extracts kinematic priors from simulation and visual priors from human observations, then fine-tunes on a small real-robot dataset. The result is a 53.7% increase in absolute success rate compared to using only real-world data and a 35.0% improvement over the best single-source pre-training baseline.

Researchers present AntiGrounding, a visual action-selection framework for robot manipulation using a dual geometric-visual trajectory interface. The framework uses a VLM to evaluate trajectory feasibility and safety, and achieves 71.25% success rate in real-world manipulation tasks. This development is relevant to AI agents as it explores multimodal reasoning and executable trajectories, with potential applications in robotics and AI-powered manipulation tasks.

Researchers introduce SafeHarness, an obstacle-aware harness for coding agents that enables safe robot manipulation. SafeHarness improves task success and collision avoidance rates by 6.5% and 27.0% respectively, surpassing the previous state-of-the-art. This development is relevant to people building or operating AI agents as it showcases an innovative solution for safe robot manipulation using coding agents.

This research proposes an onboard vision-language navigation stack for aerial robots. VLN on the Fly separates grounding, planning, and control into distinct stages, improving observability and safety. The stack uses a quantized VLM to ground instructions, a B-spline planner for trajectory generation, and a pretrained reinforcement learning policy for motor command tracking.

Researchers propose MaskHarness-WAM, a system for long-horizon robot manipulation that connects high-level task planning with low-level manipulation policies through target masks, leveraging visual feedback for subtask scheduling and execution.

Researchers have developed a method to transfer closed-loop robot policies using a coding agent that generates and improves policy code from demonstrations and simulation feedback. The results show improved success rates in policy acquisition for new tasks, with optimized references outperforming initial references in most cases.

Researchers have developed a method to improve cross-embodiment transfer in latent action models (LAMs) using action-similarity supervision. This approach allows LAMs to learn from demonstrations recorded by one robot and apply them to another robot, reducing the need for costly demonstrations. The method was evaluated on RoboTwin 2.0 and showed significant improvements in cross-embodiment transfer.

CoreSense is an architecture for robot decision-making that includes a conflict-aware belief gate to ensure auditable and safe decisions. It evaluates evidence and permits or escalates decisions. Evaluation shows reduced unsafe proceeds and improved performance in various experiments.

Researchers present HALTER, a system for automating the reset and scoring of long-horizon robot manipulation tasks. HALTER uses a graph-based approach to plan and execute resets, reducing the time spent by operators and improving the reproducibility of results.

A new framework, agentic robotic additive manufacturing (A-RAM), uses a large language model to generate execution-ready plans for robotic additive manufacturing. The framework integrates pre-execution evaluation of planning decisions and kinematic realization, improving planning outcomes and reducing jerk and extrusion times.

Researchers propose REACT, a fully spiking state-space model for event-driven temporal perception. REACT processes raw events from event cameras without temporal accumulation, allowing for fast reaction times. It achieves low-latency gesture recognition and time-to-collision estimation, with potential applications in robotic systems.

Researchers propose JEPA-WAM, a new approach to improving instruction-following in robotic manipulation by augmenting text instructions with stochastically generated visual cues. This method outperforms existing models on a real-robot benchmark, achieving success rates of 87.3%, 74.5%, and 80.9% in in-distribution, out-of-distribution scene, and out-of-distribution instruction settings, respectively.

MAGMA-GEN is a data-generation pipeline that converts ambiguous failed rollouts into validated recovery supervision for hierarchical robotic systems executing long-horizon manipulation tasks. It uses a privileged coach to hypothesize early decision-level errors and propose localized corrections or recovery actions, improving task success and recovery capabilities.

A new model, PACT-WAM, is introduced for robot manipulation tasks, using temporal context and visual foresight to predict actions and their consequences. The model reduces processing costs by using a compact temporal encoding and hierarchical history representation. It achieves high success rates on various tasks and supports a test-time enhancement through a vision-language model component.

Visual Perception Engine (VPEngine) is a modular framework for efficient GPU usage in robotic vision tasks, enabling parallel execution of multiple machine learning models. It uses a shared foundation model backbone and task-specific heads, achieving up to 3x speedup and efficient GPU utilization.

Researchers propose a method to generate force-aware manipulation trajectories for robots using a combination of video and audio generation. The approach leverages the loudness of generated contact sounds to shape a desired-force profile, enabling successful manipulation in contact-rich tasks where kinematic-only approaches fail. This development has implications for AI agents operating in robotics and manipulation tasks.