EMBODIED AI · 2026-11-04
Embodied AI — Paper Daily, 2026-11-04
Today's embodied AI track is about making robots move faster and more reliably in real deployment. W2A uses a video model only during training to teach the action model and removes it at deployment; PTS leaves the model unchanged and simply keeps the robot moving during inference, travelling faster along the planned path. JADE has the vision-language model take part in generating actions at every denoising step, and ODE Policy builds the continuity of actions directly into the policy. Reuse across robots is another thread: PointPilot predicts how objects move instead of joint actions, and Pixel designs low-cost hardware, wearable data collection and policy learning together. ReGATE lets a humanoid still pick up a box when its camera is blocked. There is also a public dataset of real-robot failures, well suited to training reward models.
4.49
When a humanoid robot reaches for a box, the box blocks its head depth camera as soon as the hands get close, yet the controller still has to finish the whole sequence of picking up, carrying and putting down. This paper lets the robot act mainly through the joint-state pathway and treats the depth image only as a correction scaled by learned gates, designed to let less of it in when the image is unreliable. Information gathered over earlier frames is kept in the memory of recurrent layers, so it is not lost the moment the image drops out. In simulation it picks up static boxes 85.2% of the time, against 48.5% for VisualMimic under the same protocol. On a real Unitree G1, after the camera saw the box once and was then completely covered, the robot still grasped the box. When the view can be blocked, relying mainly on joint states and taking in the image as needed is steadier.
4.36
In mobile-manipulation robots, the hardware, the data-collection devices, whole-body control and policy learning are usually built separately, and when they are put together they often hold each other back. This CMU paper designs all four together: a wheeled humanoid that costs under $15,000, with an upper body proportioned like a human, plus a low-cost wearable rig that lets a person record demonstrations without the robot present. Because the body proportions match, human demonstrations can teach the robot directly, and demonstrations recorded with the wearable rig alone are enough to train mobile-manipulation policies. Scored by how far each task gets completed, the full robot gets a perfect score on three ablation tasks, and disabling the torso drops the average to 47.4%, which shows the policy really uses those joints. Teams that want to build a full platform cheaply can take it as a ready starting point.
4.33
In a common kind of robot model, a vision-language model looks at the images and reads the instruction, and an action module then turns noise into an action step by step through diffusion. But the vision-language model usually just gives a fixed condition at the start and takes no part in the denoising that follows. This paper finds that letting it join every denoising step makes actions more accurate in situations never seen in training. Concretely, the vision-language model reads the current noisy action at every step and passes what it reads to the action module through a gated attention layer; during training it is also asked directly to recover clean actions from noisy ones. On four real-world bimanual tasks, average success is 20.8 percentage points higher than π0.5. Denoising together also needs fewer sampling steps, which partly offsets the cost of running the vision-language model.
4.32
Letting a robot first predict the next few video frames and then choose an action makes the policy understand physics better, but every control step then has to run a video model, which is too slow for real-time control. This paper uses the video model only during training. It first uses PCA to compress the video diffusion model's intermediate representations into a few dimensions, then adds a training loss that makes the action model's own intermediate representations match this compressed version. At deployment the video model is removed entirely. The authors' hypothesis is that although video representations are high-dimensional, the information that actually matters for actions fits in a much smaller subspace. On LIBERO it reaches 95.9% average success, with inference about 21 times faster than a representative model that predicts frames before acting. Knowledge from video pretraining can be taught to the action model during training, so inference no longer pays for it.
4.27
Every robot benchmark produces many failure recordings, yet large public datasets are almost all successful demonstrations. This paper releases all 3,718 episodes from the ArmnetBench evaluation; besides the teleoperated demonstrations, most are rollouts of 7 imitation-learning and VLA policies on 12 tabletop tasks. 73% of the rollouts are failures. Every policy for a task was trained on the same set of demonstrations and run on the same low-cost arm setup, and an operator labelled each rollout as successful, achieved with a poor finish, or failed. So on the same task one can see how many different policies fail in their own ways. Teams that want to train reward models or build failure detection finally have a sizeable, labelled set of real-robot failures.
4.07
Most robot policies output an action value directly at each control step, and whether consecutive actions connect smoothly is left for the model to learn. This paper instead learns how the action changes: a neural ordinary differential equation learns the direction in which the action moves over time, and at deployment each action is integrated from the previous one, so the executed trajectory is continuous by construction. The same situation often allows several valid moves, such as passing an obstacle on the left or the right, and learning only the average direction drives straight into it. So at the start of each chunk, a small module samples an offset from a few learned candidate directions, one per way around; every step of that chunk adds this offset to the direction of change, so the whole chunk follows one route. In imitation learning after one training epoch, its success rate is 5.9 times that of the common flow-matching policy and it generates actions 10.8 times faster; on 14 real-world tasks it has the highest success rate.
4.05
Deployed robots waste time in two ways: they stop while the model computes the next chunk of actions, and they move at the careful, slow speed of human demonstrations. This paper changes neither the model nor its training, only how actions are executed. The robot follows the path the policy planned but moves along it faster, and the path is updated only when a new observation arrives. The gripper keeps its original speed, and the speed-up is capped by the motors' capability and the inference latency. On LIBERO with injected inference latency, task throughput is 33% higher than the strongest deployable baseline for π0.5 and 104% higher for SmolVLA, with equal or better success. The losses from pausing and slow motion can be recovered without touching the model.
4.02
Different robots have different action formats, so mixed demonstration data is hard to use as one kind of supervision. This paper changes what is predicted. Instead of joint actions, it predicts how points on the task-relevant objects move in 3D space, treating the robot's gripper as one of the objects. Execution needs no trained module. Before grasping, the gripper follows the predicted motion of its own points. After grasping, the object and gripper are treated as one rigid body. A rotation plus translation is fitted from the current and predicted positions of the object's points and applied to the gripper's pose, and each robot's own controller carries it out. Because the supervision does not depend on the robot model, demonstrations from different robots can be trained together. In RoboTwin simulation, a policy trained only on ALOHA data was deployed without fine-tuning on ARX and Piper arms it had never seen. Predicting how objects move transfers across robots more easily than predicting how joints move.
3.25
GoogleA robotics team developing a policy usually switches tools between simulation, model training and real-robot deployment, moving data and models across several systems. This Google Cloud blog post introduces its physical AI offering for robot developers, which puts simulation, model training and real-world deployment on one cloud platform. It is not a new algorithm; its value is in connecting the stages: when the whole workflow runs on one platform, the turnaround from training in simulation to running on a real robot gets shorter. Whether it is worth migrating depends on a team's existing toolchain and the cost of cloud compute.
2.95
AAgiBot
In late October AgiBot released seven results over seven days, covering a dataset, a work platform, a heavy-duty robot, an agent framework, a simulation platform, a foundation model and collective experience learning. It then sent 6 G2 robots into a real convenience store for a 24-hour livestream of stocking, restocking and making hot food. Its claim is that deploying robots needs the whole technology stack working together, and a single strong model is not enough. Its simulation platform claims over 95% consistency between real-robot and simulated results on complex tasks. This is a vendor announcement with no evaluation method given; but putting robots on continuous shifts in a real store exposes long-running problems better than a one-off demo.