Robotics papers — 2026-09-22

Today we are focusing on commonsense grounded path planning because it directly addresses a gap in how robots navigate human environments. We are trying to build a system called Commonsense Ranked Search or CoRS that takes an abstract instruction like move carefully and turns it into a route that respects unstated social rules. This matters because robots working alongside people need to understand things like avoiding wet floors even if those rules are not explicitly written down in the initial programming.

The core idea is using large language models and vision-language models as a source of commonsense knowledge to reason about these hidden considerations. For example, when given an instruction and visual data, CoRS derives a specific consideration for every area, such as identifying that a wet floor is slippery and suggesting taking a detour. It then compares these considerations to rank regions by which one should avoid more, deciding if the crowd is worse than the wet floor.

This ranking sorts the areas into a commonsense order whose costs drive a standard search algorithm that always finds a valid path. This process goes beyond what recent LLM-based planners can do by discovering these unstated considerations and choosing to go around obstacles that are worth avoiding. This planning method is built on benchmarking three environments with thirteen hundred fifty problems and five instructions at various levels of abstraction, showing that CoRS successfully navigates latent constraints.

Moving down the list of significance, we have work on intelligent degradation monitoring for lithium-ion batteries which is vital for safety in electric vehicles. A new neural network framework predicts discharge incremental capacity features directly from charging signals, bypassing the need for slow, low-current measurements. The LSTM model proved to be the best choice among architectures because it balanced prediction accuracy with computational efficiency, allowing real-time integration into battery management systems.

Another area of concern is decision-aligned latent world models which aim to fix a gap where predicting a future state does not guarantee that the path taken will be successful. D-JEPA introduces a model that learns decision-relevant relations among competing futures based on executed outcomes, refining predictive geometry where action choices are most consequential. This alignment leads to improved action selection across various platforms, showing success in tasks like PushT with an eighty seven point eight nine percent success rate.

We also have advancements in making vision-language-action models run faster through native low-bit quantization using FoldQuantVLA. This post-training framework uses consistent activation representation to execute vision language action models with four bit weights and activations, achieving speedups of one point two zero to one point three three times on certain hardware. This configuration significantly boosts success rates on real robot tasks, raising observed GR00T N1.7 success from eighty percent to ninety two point five percent.

Finally, there is the work on failure recovery and monitoring for humanoid loco-manipulation through FRAMES. This framework uses a planner agent to select skills and a monitor agent based on vision-language models to evaluate skill execution against temporal observations. The monitor module showed high accuracy, detecting forty eight of fifty failures in MuJoCo trials, providing initial evidence that this failure detection component is robust.

The most critical work here is ORDER because it directly addresses the safety and reliability of knowledge-intensive robot deployments, such as those in pharmaceutical dispensing or hazardous material handling. These environments cannot tolerate the errors that models might make when faced with novel protocols.

ORDER introduces a fictitious world benchmark consisting of a three hundred forty-two thousand and sixty-nine token synthetic corpus that defines a self-consistent physics never seen in any model's training data. This corpus is paired with an ORDER-BENCH knowledge test and an ORDER-SPATIAL task requiring object ordering for safe manipulation in both familiar and entirely new scenes. GPT four point one without adaptation scored below chance on the ORDER-SPATIAL task, achieving a kendall's tau of zero point four four one, which shows its existing knowledge actively conflicts with the invented physics.

When these models undergo continual pre-training, smaller models show substantial improvement on both familiar and novel scenes. This suggests genuine world-model induction instead of mere memorization. This leads to a robot pipeline where models that pass the knowledge test often require further skill adaptation before they can produce valid plans. After this stage, small offline models outperform GPT four point one even when given retrieval access to the same rules, with a kendall's tau of zero point eight four eight versus zero point six zero six on a full perception-to-execution loop on a simulated iiwa seven arm.

This success is tied to the ORDER-SPATIAL performance rather than just knowledge test accuracy, which predicts real plan quality. This contrasts with tactile sensing research, where Tactile-JEPA shows that learning topology-aware representations using sensor connectivity graphs reduces force estimation error by six point three percent and in-hand orientation error by twenty point eight percent across various sensor types.

The most important work right now is SE-LLM-OCP because it tackles the fundamental problem of making autonomous parking safe in tricky spaces. This framework lets large language models decide on big moves, like where to park, while an optimal control module handles the fine details of keeping the car from hitting things. This combination means we can handle complex decisions without getting stuck when things get messy in a narrow spot.

The core idea is that the LLM proposes a plan by breaking it down into small steps, and then a low-level solver checks if those steps are physically possible according to vehicle dynamics. If the solver finds an issue, the LLM learns from that failure to replan better next time, which is how it automatically builds its knowledge base over time.

This approach is more significant than other work because it creates a self-improving system for a hard physical task. It means we are moving beyond just generating plausible paths to creating systems that actively learn from their mistakes in real-world constraints. This contrasts with things like HybridFlow, which focuses on making robotic actions faster by cleverly reusing network evaluations during inference, though that is crucial for real-time performance.

Another important piece of progress is HIGenNTO, which aims to generate complex humanoid interaction motions from scratch using text descriptions as a starting point. It uses optimization to synthesize these motions while respecting physical constraints like avoiding collisions and maintaining stable support, which is a big step toward creating robots that can interact naturally with humans.

The work on RiverVLN also matters because it extends vision-language navigation beyond simple indoor settings to the continuous motion required for unmanned surface vehicles navigating rivers. It introduces a phase-grounded approach that keeps track of semantic progress along the journey, which helps prevent the kind of drift seen in older navigation methods when dealing with long, ambiguous instructions.

Finally, StateMem addresses a key limitation in vision-language action policies by giving them memory so they can use information from past interactions. By using prediction errors to update a persistent memory token, it allows these policies to perform much better on manipulation tasks where context matters across multiple steps.

The work on connectivity-aware exploration is particularly important because it suggests that simply finding a successful grasp is not enough; we need to understand how those successful grasps are connected in space to explore the entire possibility landscape efficiently. This means we can learn about grasping by following structural connections rather than just trying random poses.

We saw that when we used a large-scale grasp dataset, the sets of successful grasps showed a structure that was heterogeneous and reproducible across different objects. This structure is what motivated us to introduce a connectivity-aware sampling strategy that prioritizes potential bridges between components or structural frontiers when exploring the grasp space incrementally. This method recovered the connectivity structure of successful grasp sets substantially more efficiently than random sampling or farthest-point sampling, meaning we found the underlying organization much faster.

Furthermore, we looked into whether connectivity gained from hidden grasp viability could improve subsequent grasp discovery and if experience from previously explored objects could be transferred to unseen ones. These results suggest that the spatial organization of viable actions provides information relevant to grasping beyond just the success of individual candidate actions, which motivates structure-aware exploration in robotic manipulation.

The work on visuomotor robotic pruning is particularly important because it tackles the labor intensity of maintaining modern fruit orchards through automated, closed-loop control. This method involves training an end-to-end controller entirely in simulation to guide a robotic cutter through complex branch environments to achieve precise cutpoints. It has shown success with nearly fifty percent accuracy on V-Trellis apples and forty-six percent on UFO cherries. This learning pipeline uses optical flow from a wrist camera instead of full three-dimensional reconstruction, and this learned policy has been validated in real orchards with zero-shot sim-to-real transfer, even outperforming classical methods in laboratory settings.

This success is built upon the foundation of hybrid reinforcement learning algorithms that combine offline demonstrations with online simulated rollouts to train the pruning controller. This approach addresses the challenge of learning complex manipulation tasks by leveraging both pre-existing knowledge and continuous interaction within a simulated environment.

Building on this control mechanism, another area shows how sophisticated language models can enhance human-robot interaction by enabling humanoid robots to provide intelligent, context-aware knowledge assistance through natural speech and internet knowledge retrieval. This capability moves beyond simple rule-based systems to allow for much more adaptive conversations with the MyBuddy robot.

Furthermore, advancements in perception are being made by SpectRobot, a framework that transforms sparse, high-bandwidth single-point tactile signals into image-like time-frequency spectrograms. These representations allow standard vision encoders to process tactile history while preserving crucial temporal and frequency information unavailable in conventional cameras. This technique has shown that robots can successfully solve visually occluded manipulation tasks by exploiting single-point vibration signals, and this representation can be used across different sensing technologies.

The most significant work today concerns SeeR-VLA because it directly addresses the need for robot models to possess true three dimensional spatial reasoning, which is crucial for complex manipulation. This model transforms implicit geometry from standard RGB images into explicit pointmaps that are centered around the robot, using an encoder initialized from pre-trained RGB weights to extract features and add them to existing tokens without increasing the token count. This method shows clear benefits, improving average success over RGB-only models by six point four in RoboCasa tasks and thirty two point five percentage points in real world tasks.

This improvement is further highlighted when comparing SeeR-VLA against PointVLA, where it surpasses the strongest evaluated three dimensional augmented baseline by three point five and twenty three point seven percentage points. The research also clarified how coordinate choices impact performance, revealing that centering the end effector with robot base aligned axes yields the best results. This finding is important because it suggests that when training from diverse camera setups, using robot-frame pointmaps is highly beneficial for learning.

AdaReP tackles the computational burden associated with replanning in neural world models coupled with model predictive control, which typically replans at every step to manage prediction error. By introducing AdaReP, a training-free wrapper, researchers adapt the replanning tolerance online based on current deviation and local sensitivity estimates rather than modifying the underlying world model or planner. This approach substantially reduces planner side computation while maintaining comparable task performance, achieving over eighty percent fewer queries on a fifty trial physical robot study.

InSight offers a way to achieve self guided skill acquisition for vision language action models by using a vision language model to identify missing primitives and grounding those proposals through actual robot execution. This framework allows the system to reuse known skills while collecting data for new ones without needing full human demonstrations for every task. For instance, acquired twisting and pouring skills achieved ninety two and ninety six percent success rates on hardware, compared to thirty two and sixteen percent for a zero shot baseline.

AVP introduces an end-to-end architecture that uses visual primitives to condition a flow matching action expert, aiming to simplify the process where current VLA designs entangle instruction comprehension with spatial understanding. Experiments on general pick and place tasks showed AVP improving success rates by thirty seven point zero four over pi zero point five, showing consistent gains in data efficiency and object level transfer.

CLEA proposes a closed loop embodied agent that uses four specialized large language models to manage task execution dynamically in changing environments. This framework features an interactive task planner that generates executable subtasks and a multimodal execution critic that triggers re planning when environmental perturbations exceed preset thresholds. CLEA achieved a sixty seven point three percent improvement in success rate across twelve trials, demonstrating enhanced robustness in dynamic settings.

PerchRL focuses on vision based agile perching of quadrotors on inclined platforms under rapid motion by employing a two stage strategy of state based pre training followed by vision based fine tuning. This framework uses randomized platform trajectories and temporal augmentation to improve generalization, and a hybrid learning framework during fine tuning to enhance robustness against intermittent visual loss.

HyperDet enhances 3D object detection using hyper four d radar point clouds by constructing task aware hyper four d radar point clouds through spatio temporal accumulation and foreground generative enhancement. This pipeline allows the detector to perform inference using only radar input, showing consistent improvements over standard three dimensional detectors on public datasets.

The work concerning optimal control in sizing isolated hybrid renewable energy systems matters because it directly impacts the effective carbon reduction achieved when integrating intermittent solar or wind power into off-grid setups. This research addresses the challenge of accounting for finite-horizon optimal control during the system sizing phase, which is crucial for accurately determining the necessary capacity of energy storage and renewables. The authors implemented an imitation learning approach applied to stochastic neural model predictive control to link battery storage and wind peak capacities with emissions reduction and investment costs. This method allowed them to relate these factors while respecting finite horizon optimal control without needing to solve an infeasible number of optimization problems, which is a significant computational hurdle in this area.

A case study involving an offshore system with a gas turbine, a wind farm, and a battery energy storage system demonstrated that there is a nonlinear relationship between the investment costs and the reduction in gas usage depending on the specific capacities of the wind and BESS components. This finding suggests that simply maximizing renewable penetration is not enough; the interplay between storage size and renewable capacity dictates an optimal trade-off concerning fossil fuel consumption. This sizing framework connects to other work because it establishes a way to model complex system dynamics under uncertainty, similar to how AffordanceWAM grounds visual predictions in task-relevant objects for robot manipulation.

The DiagGen framework introduces an agentic loop that takes a single image and refines it into a simulation-ready deformable asset by using a vision-language model agent to probe regions and route evidence-backed repair cues back to the generation stage. Experiments on forty assets showed these diagnostics provide useful repair cues, moderately improving the quality of the generated assets. This iterative refinement process is analogous to how HALO-WA uses latent features and action priors from world action models to adapt online, suggesting a shared theme in using learned representations for error correction in complex systems.

The work on energy-efficient air ground actuation for hybrid robots matters because it directly addresses how these complex machines can move efficiently over uneven surfaces. The research showed that by using an energy-aware reinforcement learning framework to command wheels, tilt servos, and propellers all at once without needing separate locomotion modes, the system could achieve about twenty-seven percent less mean power when navigating steps compared to the best fixed thrust allocation. This means the learned policy is significantly more efficient in terms of energy use during climbing.

Scaling vision language models with generative three-dimensional worlds is significant because it unlocks a way to train vision language action models across vastly diverse scenes without needing prohibitively expensive real-world data collection for every scenario. By using generative 3D world models alongside a language designer, the approach increased simulation success from nearly ten percent up to seventy-nine point eight percent. This method also improved real-world success from twenty-one point seven percent to seventy-five percent. This transfer capability is bolstered by the fact that increasing scene diversity directly improved zero shot generalization, showing that having more varied training data leads to better performance in unseen situations.

The SAIL framework for test time scaling in in context imitation learning is important because it shows a robust way to make imitation learning scalable by allowing robots to iteratively refine their skills using only test time compute. This method uses Monte Carlo Tree Search guided by a vision language model-based scoring mechanism, and experiments across six tasks demonstrated that increasing test time consistently improved success rates up to ninety-five percent on complex tasks. This suggests that trajectory level scaling is a solid path toward more generalizable agents.

Cognition to control introduces an object centric cognition to control architecture that preserves the physical meaning of signals between different modules. This is vital for building reliable human robot collaboration systems. This architecture achieved mean success rates between seventy-seven point one and eighty-two point one percent across nine transport scenarios when interfaces remained unchanged, significantly outperforming a scripted partner reference at fifty six point five percent. Furthermore, tests on physical unitree g1 humanoids showed a hundred percent success rate in spatially confined transport and eighty percent in handling super long objects.

Heterogeneous robot collaboration with grounded generative intelligence addresses the limitation of current teaming methods that assume well structured environments by allowing large language models to infer necessary subtasks from natural language specifications. The SPINE-HT framework, which validates subtask feasibility and updates them online using feedback, achieved nearly twice the success rate compared to prior approaches in simulation and reached an eighty seven point five percent success rate in real world missions involving a team of four robots.

SE(3) neural potential fields for trajectory planning directly from images is noteworthy because it bypasses the need for explicit 3D reconstruction by learning a field directly from poses, which removes accuracy issues associated with traditional methods. This neural potential field converged within three centimeters of the grasp on two tabletop scenes, and crucially, every path executed was collision free against ground truth geometry.

Finally, touch2robot improves the efficiency of collecting dexterous manipulation data by allowing humans to demonstrate tasks while simultaneously seeing how a target robot hand would actually contact an object. This framework reduced collection time per replay-successful demonstration from fifty eight point six seconds down to eighteen point two seconds, and policies trained on these demonstrations showed a twenty nine point one percentage point improvement over visual only feedback.

Today's papers

The papers

Important terms

Commonsense Ranked Search (CoRS)
A planning method that uses LLMs and vision-language models to derive unstated social rules from instructions and visuals. It ranks areas based on these hidden constraints to find a valid path that respects common sense, like avoiding wet floors.
Decision-aligned Latent World Models (D-JEPA)
A model designed to fix the gap where predicting a future state doesn't guarantee a successful path. It learns decision-relevant relations among competing futures based on past outcomes, improving action selection for tasks like PushT.
SeeR-VLA
A vision-language model that transforms implicit geometry from standard images into explicit pointmaps centered around the robot. This allows for true 3D spatial reasoning, significantly boosting success in complex manipulation tasks.
ORDER framework
A benchmark and model approach testing knowledge robustness by creating a synthetic world with physics unseen in training data. It shows how models can induce genuine world-model knowledge through continual pre-training.