Three days building a robot at the UK RAS STEPS MicroMaze Robot Hackathon involved constructing a small autonomous robot to navigate increasingly difficult mazes. Participants utilized components like sensors, motors, and a Raspberry Pi Pico W while debugging the system. Beyond the technical challenges, the experience fostered strong friendships among the team members. The event also provided a valuable opportunity for networking and shared experiences in York. Ultimately, the process highlighted that the most valuable outcomes are often the unexpected relationships forged during the building process.
Boston Dynamics has appointed Rohit Prasad as CEO, effective October 7, to drive the commercialization of its robotic systems. Prasad is a former Amazon AI leader who contributed to Alexa and the Nova foundation models. The appointment emphasizes integrating advanced AI into the company's robotics portfolio, which includes Spot, Stretch, and Atlas. Prasad's tenure followed Robert Playter's retirement, which followed a period of transition. The move aims to bring advanced AI capabilities to industrial and commercial environments.
MIT researchers developed "SANDO," a method to plan flight paths for UAVs that avoid unknown, moving obstacles. This system mathematically guarantees collision-free routes through unmapped environments, even when obstacles move unpredictably. SANDO creates a time-sensitive safety corridor by considering the future movements of dynamic obstacles based on their maximum velocity. The planner uses heat-map techniques and optimizes trajectories to ensure the fastest, safest path while allowing for rapid onboard replanning. This approach offers formal safety guarantees essential for high-stakes applications like search and rescue missions.
This work proposes an explainability-guided multi-objective evolutionary approach for robust testing of DNN-based robotic navigation systems. The method generates sparse perturbations over representative images using visual and behavioral clustering, guided by aggregated Integrated Gradients. Evaluation was performed on a DNN-controlled robot in Gazebo and validated on a physical robot. The approach yielded a 70.0% median success rate, significantly outperforming unguided search methods. Simulation results correlated well with physical outcomes, though physical validation remains essential.
Riemannian Barrier Metric RRT (RMRRT) is a motion planning framework unifying equality and inequality constraints for sampling-based planning in high-dimensional robotic systems. It constructs a unified local geometry by building an ambient barrier metric from inequality terms and inducing a tangent-space metric via a (G)-orthogonal projection for equality constraints. This tangent-space metric is consistently used for steering and nearest-neighbor selection, biasing exploration away from inequality boundaries while maintaining equality consistency. Experimental results show RMRRT achieves a 100% success rate in constrained tasks and reduces planning time compared to baselines. Ablation studies confirm the metric improves exploration quality by reducing rejected samples and path length.
The study evaluates whether a learned corrector outperforms a simple retreat when recovering a frozen vision-language-action (VLA) policy. This comparison is done across four RoboTwin tasks using frozen $\pi_{0.5}$. The learned pipeline showed a significant success increase of $+13.5$\,pp on the `beat_allowbreak` task, but no gain on the other three tasks. Furthermore, while a fixed-time trigger produced a net gain, this difference was not consistently met with the learned pipeline. The decision to intervene is task-dependent, necessitating paired baselines for meaningful evaluation.
Vision-Language-Action (VLA) models often use a single policy, which is insufficient for varying task states. SWAP introduces StepWise Action Policy Routing to dynamically compose multiple VLA policies during execution. This is achieved by framing policy routing as an offline reinforcement learning problem to learn a routing critic. SWAP allows robots to select new policies online instead of committing to one for an entire episode. Evaluation shows SWAP improves real-world task success by up to 33% and reduces successful action step lengths by 28.3%.
A new phantom device is presented to facilitate real-world Reinforcement Learning (RL) testing for physically interactive algorithms. This device replicates shoulder ball-and-socket motion along with scapular and protraction/retraction movements. It allows for the demonstration of multiple 3D joint limits in physical experiments. A Franka Panda arm was used to interact with the phantom as an arm for activities of daily living and physical human-robot interaction. This work addresses challenges related to repeatability and safety in human-robot physical experiments.
Object-goal navigation in multi-floor scenarios is challenging due to sparse rewards in long-horizon decision-making. A modular framework is proposed to diagnose failure factors using an effective learnable policy. This involves factorizing the global policy into intra-floor exploration and inter-floor switching policies. The intra-floor policy is initialized by distilling Visual Language Model exploration logic for Reinforcement Learning. Experimental results suggest that perception performance and stair climbing stability are the main bottlenecks in this navigation task.
ACG-WAM models world actions by jointly learning visual prediction and robot actions using scene evolution observations for policy learning. It introduces an auxiliary objective, ACG-JEPA, to predict geometric features across time horizons from current observations and actions. This prediction is achieved by using the future slot of an Afrozen VGGT encoding of image pairs as the target. The method successfully achieves high success rates in simulated environments and shows significant improvement over existing methods on real robot tasks. ACG-WAM achieved 93.46% success in clean scenes and 85.00% success on real robots.
ProactiveVLA augments embodied memory by using proactive environment exploration to acquire reusable knowledge for deployment-time adaptation. The system combines a reasoning agent with a frozen vision-language-action model (VLA) to adapt through execution feedback and memory. After an initial task, the agent focuses remaining interaction on self-proposed goals concerning object affordances and state changes. It consolidates this task-directed and exploratory experience into memory to guide future planning and control. ProactiveVLA demonstrated superior performance over baselines on benchmarks like LIBERO-Pro.