Robots are learning to move smoothly, feel, and recover from failure

A multi-agent robot that recovers

A preprint introduces A2-MEAS, an embodied multi-agent framework that uses learned correction codes and a vision-language-action model to enhance robot control and failure recovery. The system achieved a 0.6411 Q-score on the 100-task BEHAVIOR 2026 public subset, and a 0.5755 Q-score on the 50-task subset, using a vision-language-action model with motion-aware history and latent world modeling. The framework uses code as a scaling interface, the authors say, combining corrective language, spatial evidence and execution history to improve robot autonomy. The team plans to release the models, training scripts and infrastructure. m:…7740

The real-robot numbers are the ones that matter most. A2-WLA achieved 57.92% average success across 12 real-robot household tasks, compared with 26.67% for the pi0.5 reference, and the complete A2-MEAS system achieved 95.05% average success on RoboTwin 2.0. Training infrastructure for A2-MEAS delivers 427.1 samples per second, which the authors say is 4.03 times the throughput of a pi0.5 reference implementation. The gap between simulation and the real world is still large, 57.92% is not a product, but it is a substantial improvement over the leading baseline on the same tasks, and the correction-code mechanism offers a reason why: the system can notice a failure and correct it, rather than starting over. m:…7740

Smooth motion by construction

ODE Policy takes a different approach to the same problem, changing what the policy learns. Traditional robot policies map observations to discrete action values, but ODE Policy learns the vector field of the action using neural ordinary differential equations, matching integrated actions to targets over short action chunks in imitation learning and over single control steps in reinforcement learning. Executed trajectories are continuous by construction, the authors say, and a diversity head captures multi-modal behavior by randomizing the direction of the field. The method is designed to function as a visuomotor policy, an action head for vision-language-action models, and an actor in model-based reinforcement learning. m:…7741

The reported gains are large. In imitation learning, ODE Policy reaches 5.9 times the success of flow matching after one epoch, with 5.4 times the generalization, and generates actions 10.8 times faster than flow matching. It achieved the highest success rate on 14 real-world tasks, tested against 14 baselines, across unimanual, bimanual, dexterous and humanoid robotics, and produces smoother motion than human demonstrations, the authors say. Smoothness matters for reasons beyond aesthetics: jerky motion wears hardware, drops objects and makes a robot unsafe to work near. A policy that is smooth by construction removes a whole class of failure rather than patching it. m:…7741

Cheaper eyes, shared touch

FOVI-Dex addresses the cost of robot perception by introducing foveated rendering and perception to optimize robotic reinforcement learning for dexterous manipulation. The method uses 8.9 times fewer vision tokens than high-resolution fixed-camera control, supports 6 times more parallel environments, and achieves 4.4 times higher training throughput. Training is performed via online reinforcement learning without demonstrations, and the system is a language-conditioned visuotactile policy. It outperforms high-resolution baselines and matched uniform active camera baselines in both simulation and the real world, the authors say. m:…6147

LingBot-Touch solves a different bottleneck: touch. It is a tactile foundation model that creates a unified, extensible representation for diverse tactile sensors, including piezoresistive, magnetic and visuotactile types, by generating plug-and-play tactile tokens that integrate into existing manipulation policies such as diffusion models and vision-language-action architectures. The approach aims to overcome sensor-specific physics limitations, the authors say, and experimental results in simulation and real-robot deployments indicate that the representation enhances manipulation dexterity while maintaining computational efficiency. A shared touch representation means a policy trained on one sensor can, in principle, work with another, which is how dexterous manipulation starts to scale across hardware. m:…7625

What is still missing

The honest reading of these four preprints is that embodied AI is improving fast on the measures that researchers control, and still far from reliable on the measures that customers control. A 57.92% success rate on household tasks is a research result, not a product, and the highest success rate on 14 real-world tasks is a comparative claim, not an absolute one. The advances are real, though, and they share a shape: smoother action generation, cheaper perception, shared touch and explicit failure recovery are all ways of making a robot do more with the same data and hardware. m:…7740 m:…7741

For a robotics founder, the practical implication is to invest in the training and perception infrastructure, not just in demonstrations. A2-MEAS training infrastructure delivers 427.1 samples per second, and FOVI-Dex supports 6 times more parallel environments than the high-resolution baseline, which means the teams with better infrastructure can iterate faster on the same ideas. The next question is whether these methods compose: a smooth ODE policy, with foveated perception, shared touch tokens and correction codes for failure recovery, running on one robot. None of these papers tests that combination, and that is where the next advance is likely to come from. m:…7740 m:…6147