HexaShell is a lightweight framework that lets coding agents control a dual-arm robot via the command line. It achieved a 39.7% success rate and a 48.0 progress score on all 42 RoboDojo tasks, a new best result for RoboDojo.
The research team achieved Baoding ball rotation on the Shadow Dexterous Hand Lite using only proprioception and binary tactile feedback, enabling zero-shot sim-to-real transfer. In a single continuous trial on the physical hand, it completed 225 180° rotations (112.5 full rotations) at about 0.36 rotations per second, with no vision or object-state information throughout. The authors also compared the roles of touch and pure proprioception, identifying the hand regions where tactile information is most critical.
AWAM (Active World-Action Model) introduces a virtual active-perception paradigm that actively plans observation positions and, before execution, imagines what will be seen along the planned trajectory. It treats the predicted task-related virtual camera motion as a structured query into a geometry-conditioned world model, yielding a reusable spatial prior that is compressed into a fixed memory and fused with live observations through cross-attention for diffusion-based action prediction.
The paper proposes ActiveAlign, a data adaptation framework that addresses the mismatch between the initial observation state produced by active camera motion and the manipulation policy's training distribution (first-frame distribution sensitivity): on LIBERO, directly prepending camera repositioning to pi_0.5 reduces the success rate from 46.3% to 11.5%.
The paper proposes an Interventional Memory Evaluation Protocol (IMEP) for vision-language-action (VLA) systems, used to determine whether memory truly influences robot decision-making, covering four questions: the net benefit of memory, the conditions under which it takes effect, changes to task outcomes, and whether the effects are specific to task-relevant facts.
The authors propose LAW, a closed-loop VLM-WAM framework that unifies language reasoning, action generation, and world modeling through semantic subtask progress, achieving an average success rate of 94.21% across 50 RoboTwin tasks.
arXiv has published a survey that systematically reviews the memory mechanisms of long-horizon robot Vision-Language-Action (VLA) models, dividing them by memory scope and persistence into five representative paradigms and one emerging frontier. The survey compares representative VLA memory systems across memory characteristics, architectural complexity, and open-source maturity, finding a trade-off in which greater architectural complexity is generally accompanied by lower open-source maturity. It also discusses key challenges in building full-lifecycle memory systems, including efficient memory management and lifelong knowledge evolution.
CoEvolve is a framework that represents team coordination as executable code over fixed robot motor skills. Through verifier-guided co-evolution, multimodal foundation models instantiate coordination strategies, physical verification evaluates each executed transition, and rejected behaviors trigger structured revisions starting from an accepted checkpoint.
LY-GWM-RoboCasa-Human300 combines a Human300-fine-tuned GR00T-N1.5 action generator, an action-conditioned graph dynamics model, and a binary state-reward scorer; at inference it samples 4 action chunks and executes the candidate with the highest predicted-state reward.
The NeurIPS 2026 paper NestedVLA proposes using a nested hypernetwork to generate task-adaptive action-expert parameters for a target task, combining them with a frozen base VLA. This replaces parameter aggregation with parameter generation, mitigating interference in multi-task training.
ORD-WM proposes a 3D world model that learns object-centric rigid-body dynamics for robotic manipulation tasks. It uses a vector-neuron variational autoencoder supervised by signed distance fields to encode each object point cloud into a geometry-aware latent representation composed of 3D vector features, and uses an action-conditioned dynamics model to predict per-object SE(3) motion increments. It then analytically applies the predicted rigid-body transformations to the current latent state rather than generating an unconstrained next latent state, thereby separating persistent geometry from pose changes.
SCOPE (Selective Certification Of Program Edits) uses failed and successful executions as complementary evidence to guide program revision: frontier-guided program search annotates a shared diagnostic graph with execution evidence to distinguish verified progress from unresolved checks, while selective behavioral certification uses stateful API replay to restore recorded responses and mutations and checks whether edits preserve selected historical successful behaviors.
The paper proposes BeJEPA, a goal-conditioned humanoid behavior world model learned through unsupervised reinforcement learning from human motion priors and robot interaction, using a joint-embedding predictive architecture (JEPA) to learn a shared representation of body configuration and world placement.
GRASP proposes a training-phase prior injection framework for bimanual robotic manipulation. A physics-grounded flow-matching loss uses differentiable forward kinematics (FK) to supervise tool center point (TCP) poses and inter-arm relative poses, and the framework introduces a global query group distilled from DINOv3 together with local interaction queries supervised by Grounding DINO; the teacher models are removed at inference. Simulation and real-world long-horizon bimanual task experiments show that the two priors are complementary under different deployment conditions: the full model performs better in in-distribution deployment, while the distilled query branch is more suitable under strong visual randomization.
InstAgent proposes a recursively self-improving (RSI) agent operating system that uses agentic behavior trees (BTs) to explicitly model capability assignment, inter-robot dependencies, and handoff checks for heterogeneous robots, and updates collaboration programs from execution feedback while preserving successful revisions for reuse in later tasks. Experiments on 20 real-world multi-robot tasks with 3 high-level models show consistent gains in task success rate, first-attempt completion rate, and execution efficiency.
A critical arXiv review systematically examines the architectures, physical adaptation, and validation pathways of foundation models in autonomous space robotics, focusing on orbital manipulation and proximity operations. The article notes that event cameras can provide timely motion observations under strong illumination contrast, that spiking neural networks offer a possible route to sparse local computation, and organizes validation as an iterative "simulation-ground-space" closed loop, using a satellite capture case to connect observation, approach, contact, and recovery. The full text is 23 pages with 7 figures, accompanied by a source list and evidence document.
The paper proposes an architecture-agnostic probabilistic formal framework based on stochastic kernels for uniformly describing embodied foundation models. The framework has two parts: characterizing models with conditional output distributions and action-realization kernels to induce task-conditioned robot policies, and connecting pretraining, post-training, and policy deployment. The authors instantiate it on OpenVLA and DreamZero, covering both autoregressive action-token prediction and joint video-action generation paradigms.
VLA-Fabric assigns a complete vision-language-action (VLA) model to each robotic arm, preserving each arm's local perception and action ownership, and achieves coordination through explicitly learned interaction across policy boundaries. The paper systematically studies where interaction should occur, how peer information should be aggregated, and how interaction should be learned, finding that shared context, peer-specific latent variables, and action-stage interaction play complementary coordination roles, and reproduces this principle on structurally different VLA architectures.
The paper proposes the MAP (Masked Action Space) interface, which unifies spatiotemporal references such as keypoints, flows, orientations, and key poses as partial constraints in the same temporal Cartesian action space, and models controllable manipulation as an action inpainting problem.
MambaAct proposes replacing the Transformer-based action expert in π0.5 with a compact Mamba action expert while retaining its pretrained vision-language prefix encoder.
A CoRL 2026 paper proposes SpatialBEV, which queries bird's-eye-view (BEV) features around the predicted state at each step and conditions the next prediction on current geometric and semantic cues, in order to capture the complex interactions between terrain and vehicle in off-road scenes.
An arXiv paper proposes AnchorQ, which uses flow reversal to map reference policy behaviors into state-conditioned latent anchors, samples particles near the anchors, and applies value-guided sampling to achieve coarse-to-fine policy improvement. The analysis shows that anchors reduce identification complexity but introduce support bias; particle refinement trades this bias against action-class complexity, and critic accuracy determines the reliability of KL-regularized value-guided selection.
DefFlow is a goal-conditioned manipulation framework for deformable objects. By jointly predicting the goal-conditioned 3D point flows of the robot and the deformable object from demonstrations, it captures manipulation decisions such as grasping, releasing, and re-grasping, as well as the continuous fine motions required to reach the goal. Across rope, granular media, and cloth settings with diverse initial and target configurations, DefFlow clearly outperforms goal-conditioned action prediction and robot-flow-only baselines on success rate and goal-distance metrics, and it was deployed on a real robot using real point cloud observations and unseen target configurations.
This work proposes a visuomotor imitation learning pipeline for pedicel-targeted strawberry harvesting that unifies perception, motion planning, and contact manipulation: data is collected from human teleoperation demonstrations, an Action Chunking with Transformers (ACT) policy with a ResNet-50 visual backbone is trained, and the policy is deployed autonomously on a Jetson AGX Orin edge platform at 15 Hz.
FCTS is a failure-conditioned terrain curriculum framework that uses localized traversal failures to automatically generate the terrain a policy should train on next. Each failure's location, termination mode, robot state, and surrounding terrain geometry jointly condition the generation process, and Monte Carlo Tree Search explores alternative curriculum trajectories under a fixed training budget.
The paper proposes Embodiment-Classifier Guidance, which trains a lightweight embodiment classifier and uses the classifier's gradient during denoising steps to guide a frozen π0 policy, enabling transfer to new robot embodiments without additional demonstrations or parameter updates.
An arXiv paper proposes RePlanMem-VLA (Replanning-Guided Keyframe Memory for VLA), which turns a policy's own plan revisions into an online memory-write signal: during receding-horizon execution, if consecutive action chunks show unusually large disagreement in their predictions of the same future action, this is treated as a task-relevant event.
The paper proposes RAX (Robot Agent eXploration), an open-source robot agent framework centered on scaling online exploration, in which a general-purpose coding agent generates robot control code without using action demonstration data, hand-engineered action primitives, or value-function guidance.
Why it matters: The paper compares the imitation learning paradigm that relies on demonstration data with a demonstration-free online exploration approach, and reports quantitative gaps in simulation and on real robots, helping readers assess the feasibility of this direction.
The SenseNova joint research team released a report evaluating GPT-6-Astra's embodied capabilities in robotic manipulation and navigation. Under a unified protocol, the team evaluated it on 12 representative benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands, allowing in-context learning (ICL) from training-set demonstrations and substantially increasing the number of execution steps, achieving what the paper calls unprecedented performance.
A search for the keyword "robotics" on OpenAI's official website returns 27 open positions, up from 11 in May this year. The roles span algorithms, control, actuators, sensors, firmware, data collection, manufacturing, supply chain, and other areas. The recruitment description for supply chain positions states, "We're a growing hardware company."
Seerbotics demonstrated its self-developed VLA model driving a robotic arm to complete the seven-move sequence of a three-disc Tower of Hanoi. When an observer slides the base to change its overall position or rearranges the discs into an arbitrary intermediate state, the robotic arm can still re-identify the current configuration, determine the next move, and continue completing the remaining task. Trained on only 60 demonstration trajectories, the model can generalize across changes in object position, color, and shape. The Seerbotics robot "brain" consists of three layers—model, cerebellum, and nerve—responsible respectively for scene understanding and decision-making, motion planning and execution, and a unified interface connecting sensors and joint actuators.
The paper proposes Ropedia W0, an egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. The research builds the Xp-Data dataset, containing 10,000 hours of egocentric human data and 3,000 hours of robot data, with unified multimodal annotations and quality control; camera poses encode locomotion and projected skeletons encode manipulation, with robot skeletons combining each embodiment's forward kinematics and manual keypoint annotations.
The paper proposes Anti-Persistence, a memory design that accumulates trajectories and outcomes, reconstructs task guidance on demand, and discards it after execution, so that errors in stored abstractions are not repeatedly reused. The authors call the phenomenon in which errors are regenerated through a cycle of abstraction, execution, and re-abstraction error reproduction. Across multiple benchmarks and backbone models, Anti-Persistence achieves higher held-out task success rates and the lowest average LLM call cost among the evaluated memory designs.
Westlake Robotics has upgraded its general-purpose brain WR1, adopting a dual pretraining architecture of a general-purpose large brain and cerebellum, integrating a world model with VLA to generate action sequences, and executing full-body control through its self-developed general action model GAE. In the demonstration, WR1 completed a continuous task sequence: taking clothes from a drying rack, folding pants with two machines working together, and across the kitchen, living room, and bedroom, fetching corn, storing dolls, and tidying a quilt. The system introduces long- and short-term motion state review and a real-time action chunking mechanism to reduce pauses and abrupt changes in long-horizon tasks.
UC Davis and ADI have open-sourced Agile-WAM, a tactile world action model for contact-rich manipulation. Through multi-scale, multimodal prediction, it lets vision be supervised over long spans and touch over next-frame spans, achieving a 29.4% relative improvement in overall success rate across five real-robot tasks. On a single RTX 4090, inference latency is 10.35 ms with a control frequency of 96.62 Hz, about 40 times higher than the diffusion policy DP; it achieves best or near-best results on multiple of 9 ManiFeel simulation tasks and Flexiv Rizon 4 real-robot tasks.
VAST has released and open-sourced its real-time world model EdenLive, opening up the model weights, inference code, technical report, and an end-to-end demo. According to VAST, on a single consumer-grade Nvidia RTX 5090 GPU at 720P, the 5B model generates at over 24 FPS, can generate continuously for minutes at a time, and has end-to-end latency in the hundreds of milliseconds.
On October 16, the China Securities Regulatory Commission issued a filing notice to Hefei Youibot Robotics Co., Ltd. for its overseas offering and listing and the "full circulation" of its domestic unlisted shares. The company plans to issue no more than 49,477,800 ordinary shares for overseas listing on the Hong Kong Stock Exchange. On October 12, Jiangte Electric announced a plan to contribute 100 million yuan to subscribe for newly added registered capital in Vita Power. After the capital increase is completed, it will hold a 3.0303% equity stake, with a total contribution of 300 million yuan together with related investors, and the parties have already carried out joint development of dexterous-hand motors and key components.
Hiwonder's Tonybot and TonyPi Pro are two humanoid robot platforms aimed at AI learning. The former is based on ESP32 with 17 degrees of freedom, focusing on low-barrier entry via Scratch/Arduino/Python and on servo and sensor teaching; the latter is based on Raspberry Pi 4B or 5 with 20 degrees of freedom, equipped with a 2-DOF head vision gimbal and openable claws, and is aimed at OpenCV and MediaPipe vision and embodied scenario demonstrations.
A team at HKUST-GZ proposes UniWAM, which uses a Mixture-of-Transformers (MoT) architecture to integrate physical reasoning, future visual generation, and action prediction into one model. Built on three expert modules—Qwen3-VL-2B-Instruct, the Wan2.2-TI2V-5B backbone, and an action predictor—and paired with the EgoANT pipeline for automatic annotation of egocentric human video, it is jointly pretrained on human data, robot demonstrations, and VQA data.
RoboTTT targets a structural flaw in robot foundation models: mainstream policies such as VLA struggle to remember things that happened earlier. It proposes using fast weights and context scaling to improve long-horizon memory. The original text is an excerpt from an introduction and does not provide full experimental details.