EMBODIED AI · 2026-11-16
Embodied AI — Paper Daily, 2026-11-16
Today's embodied track is about how robots move in real time. One study finds that a VLA's reactivity depends heavily on how it is executed, and sometimes changing only the execution horizon helps as much as retraining. XInfer lets one GPU queue requests by how many actions each robot has left, so it serves more robots. HDH trains a cheap action judge from the robot's own record of successes and failures. ILA and FORGE make humanoids more compliant and stable when pushed, pulled, or carrying things. SALA and SOMA Retargeter both look for ways to turn human motion into supervision robots can use. And six AGIBOT robots staffed a convenience store for a full day.
4.75
A VLA model predicts a short chunk of future actions at a time, and when objects move, that chunk goes stale before it finishes executing. The authors use a pretrained π0.5 on two dynamic simulation benchmarks and three real-robot tasks. They systematically compare how far to predict, how many steps to execute each time, and synchronous or asynchronous inference.
4.70
The authors' company already has hundreds of VLA-driven robots working in real environments, and robots must share GPUs to keep inference costs down. Robots keep executing old actions while waiting for new ones. So among requests of the same size, some robots are about to run out and need a reply now, while others can wait.
4.50
The actions a VLA model samples are not always good, and sampling several and picking one is safer, but using a large model as the judge adds a long wait to every query. The authors' HDH trains its judge only on records from the frozen VLA's own deployments. The policy first runs a batch of tasks on its own and the final outcome after each action chunk is recorded. A small model then learns to predict, from the current view and the actions already executed, how likely a candidate chunk is to end in success. At deployment the VLA generates two candidates and the small model executes the higher-scoring one, with no expert labels and no predicted future frames. On four LIBERO-10 tasks, average success rises from 78.8% to 82.3% with only 5.4% more query latency, more accurate and faster than GPT-family judges, and the gain is larger on three real-robot tasks. A robot's own record of successes and failures is enough to teach a cheap action judge
4.49
Humanoid robots that track human motion are often used for teleoperation and motion data collection. But following the reference motion rigidly is unsafe in tight spaces or when a person is in physical contact, and it struggles to yield when someone guides its arm. ILA leaves the original tracking policy unchanged and adds an adapter in front of it: from the robot's recent observation history, it infers where an external force is trying to move the motion and edits the reference handed to the tracker accordingly. The adapter is trained entirely in simulation and needs no dedicated force sensor. The same trained adapter plugs straight into three independently trained trackers with neither side retrained, cutting the error of guiding an arm to a target by 63.6% to 75.4% while largely keeping normal tracking. On hardware, none of the original trackers could be pulled to a lateral-raise target, while every tracker with the adapter reached it in all ten trials. Compliance can be a generic plug-in layer outside the tracking policy
4.47
Robot interaction data is expensive, while first-person human video is cheap and plentiful, and a common approach uses latent action models to extract action signals from video to supervise policies. The problem is that most of these models learn only by reconstructing video, so their latent actions cluster by how the scene looks rather than by what action was taken, and human actions fail to line up with robot actions. SALA needs no paired human and robot data. It uses the text descriptions of unpaired human and robot clips to estimate semantic similarity, pulls human latent actions toward semantically similar robot latents, and grounds the latent space in control through robot action reconstruction. Human video then helps in two directions. More general human video in pre-training keeps lowering the robot's action prediction error on downstream tasks. With teleoperation data fixed, adding human demonstrations of the target task in post-training raises average real-robot success from 53.75% to 72.5%. Latent actions must align by action meaning before human video can truly supervise robots
4.46
Getting a humanoid robot to walk while carrying things makes teleoperated data collection hard: the controller must be precise, stay balanced, and know how much force its hands feel. FORGE gives the operator coarse-to-fine commands for walking speed, where the feet land, where the wrists go, and an optional desired force, usable alone or together. Training needs no human whole-body motion trajectories, and deployment needs only the two wrist poses relative to the pelvis plus rough movement commands. With no force sensor at the wrist, it estimates hand forces from joint states and low-level motor torques, adding a simulator-supervised learned correction on top of a momentum-based estimator. A frozen walking policy anchors the lower body, and a residual policy learns the upper body plus any needed lower-body corrections, which makes training more stable. A diffusion policy trained on top of it carries boxes from 0 to 2 kg on the real robot, 20 trials per load, with at least 18 successes even on the hardest one. Without wrist force sensors, motor torques can still give a usable force estimate
4.45
When a person teleoperates a robot, can a large model tell what they want just from the video? The authors split this into two levels, the action happening now and the overall task it serves, and build a dataset of teleoperation episodes from many kinds of robots. Off-the-shelf large models do poorly here. VIM, fine-tuned on this data, raises the semantic similarity between predicted and true task descriptions from GPT-4o's 0.470 to 0.821. The more practical result is downstream: the task VIM infers is handed to a VLA policy that takes over midway and finishes the remaining manipulation, and in unseen real environments the human teleoperation time drops 34.6% compared with using GPT-4o to infer intent. Guessing the operator's intent correctly lets the robot take over midway
3.77
NVIDIAIn robot pick-and-place simulation, a robot can look correct while stationary and still move or touch objects incorrectly once the simulation runs. This NVIDIA tutorial builds a beverage-can pick example with the SimReady asset library on Hugging Face. SimReady means an asset already has the physical properties needed for a given simulation use, such as collisions, mass, joints, and joint drives. The tutorial places an arm, a gripper, and a can in Isaac Sim, then has an AI agent that can call Isaac Sim skills assemble the robot, check joints and contacts, and generate control code to grasp, lift, move, and release. The authors also state the limits: the can starts at a fixed position, no vision is used, collision checks are limited, and the can neither deforms nor simulates liquid, so the result says nothing about uncertain positions, other payloads, or real hardware. Assets with the right physical properties are only a start, and the assembled scene still has to be run to know