EMBODIED AI · 2026-10-27
Embodied AI — Paper Daily, 2026-10-27
Embodied AI has 10 papers today, one with a deep dive. The deep-dive paper finds that the longer a flow-matching VLA (vision-language-action model) executes blind, the fewer numerical integration steps it should use when generating actions. On long-horizon simulation tasks π0 goes from 86.9% to 90.5% success while inference gets several times faster, and the effect holds on a real robot. Several others cover robot data: OXE-AugE extends the OXE dataset to more robot embodiments, NVIDIA open-sources a robot data labeling engine, and one paper uses reinforcement learning to generate more efficient and diverse demonstrations. Teams deploying VLAs can start with the first paper.
4.75
In flow-matching vision-language-action models, solver steps and execution horizon turn out to be coupled, and the best setting can reverse. These models start from random noise and use a numerical solver to push it step by step into a smooth action chunk. Deployment usually tunes the two separately: the horizon amortizes inference latency, while solver steps stay high for accuracy. Under frequent replanning, fewer solver steps do hurt, but when a long stretch of actions runs open loop, fewer solver steps actually help, because coarse integration keeps actions closer to the mean trajectory and limits compounding drift. Executing 5 actions at a time, 10 solver steps win. Executing 20 at a time, 2 solver steps win. The pattern holds across benchmarks and real robots, cutting latency and raising success without training.
4.77
Generalist robot policies must control many arm and gripper combinations, but recollecting data and retraining for each new robot is too costly. The widely used Open X-Embodiment dataset gathers over sixty robot datasets yet is badly imbalanced: the top four robot types make up over 85% of its real data. This paper builds an augmentation toolkit that extends existing demonstrations to nine other arm and gripper combinations, yielding OXE-AugE with 4.4 million trajectories, more than three times the original. The abstract gives the scale and findings but not how actions are converted to the new robots. Experiments show diverse robot augmentation helps not only the augmented robots but also unseen robots and the original robots under distribution shift.
4.75
NVIDIATraining robot foundation models takes huge amounts of labels: deployments need task-success judgments, reinforcement-learning post-training needs per-step process scores, and pretraining needs timestamped captions to filter data. NVIDIA has open-sourced Flora, the data engine used to train GR00T, where one model produces all three kinds of labels. It sets the best scores on public captioning and reward benchmarks and matches GPT-6 Astra on internal tests. It is roughly 125 times cheaper and labels 200,000 hours of data a day on 512 GB300s.
4.52
Imitation learning depends heavily on how widely its demonstrations cover the task, but collecting diverse successful demonstrations is expensive. This paper uses reinforcement learning as a data generator rather than a deployable controller. DiscoDemo trains a generator where each latent code yields a different way to solve the task; training starts from states near the goal and gradually backs off. A diversity reward grows as the states visited under different latent codes lie farther apart in a learned representation, and it is given only to trajectories that complete the task. The generator's trajectories are filtered: failures are dropped, then any with unwanted robot-scene contact or excessive insertion force; the rest become a static dataset for a downstream visuomotor policy. Across four manipulation tasks, average downstream success versus the strongest baseline rises from 40% to 62% in simulation. On the real robot it rises from 39% to 78%.
4.50
Vision-language-action models are too large for the edge devices robots carry. Rather than compressing after training, BitVLA uses extreme low-bit weights from the start: every parameter is one of minus one, zero or one, about 1.58 bits, conventionally called a 1-bit model, built on the open BitNet LLM. The vision encoder's weights are quantized to the same three values, and training pulls its image representations toward those of the original full-precision encoder. It matches the full-precision OpenVLA-OFT baseline while cutting model memory by 11 times and end-to-end latency by 4.4 times.
4.49
People can spin Baoding balls in one palm without looking, but robots struggle to manipulate for long by touch alone. Blind policies can be trained in simulation, yet on a real dexterous hand contacts change constantly, tactile signals drift with motion, and simulated dynamics differ from reality. This paper uses only joint sensing and on-off tactile signals on a Shadow dexterous hand. In simulation it reproduces the real tactile sensor layout and the tendon-driven coupling of finger joints. It calibrates the on-off threshold of each tactile sensor separately, and adds two new randomizations: random limits on how fast actions can change, and randomly stuck tactile sensors. The policy transfers zero-shot to the real hand, completing 225 consecutive half-turns in one continuous trial with no vision or object position.
4.49
Dexterous hand demonstrations must keep fine hand-object contact while staying executable at deployment. Existing pipelines either lose fine motions when converting human hand movement to robot hand movement, or rely on costly, unintuitive robot-specific teleoperation rigs. RealDexUMI is a wearable collection device that uses the same dexterous hand module for collection and deployment, with in-hand cameras and fingertip touch sensors. A palm-side glove whose joints mirror the robot hand one to one, so bending a finger joint bends the matching robot joint with no motion conversion. Views, touch and actions match exactly between collection and deployment. Across eight real-robot tasks, average success is 88.75%, and policies transfer across three robots.
4.47
When large robot foundation models are adapted to new tasks with offline teleoperated demonstrations, the data rarely covers the hard states of deployment, such as contact, recovery from mistakes and precise alignment. TIDE puts robot execution, human takeover, learning and policy updates into one deployment session. The system records whether each action chunk came from the human or the policy, and training feeds that source label to the model as an extra input. When the robot runs on its own, the label is fixed to the human source, so the policy acts the way human takeovers did. No value function, reward model or action scorer is needed, and updated policies go onto the robot at the end of each episode. On four real bimanual tasks, task progress and full completion beat both the starting policy and extra offline demonstrations, and the share of human takeovers falls.
4.47
In theory, regularized reinforcement learning has provable robustness and should ease moving policies from simulation to real robots, yet pipelines still default to the weakly regularized PPO. This paper systematically compares common algorithms on sim-to-real transfer and derives a new one, SMM-AC. Instead of taking the maximum over action values when estimating the next state's value, it uses a smooth maximum: multiply the action values by a parameter ω, exponentiate, average, take the log and divide by ω. A large ω approaches the maximum and ω near 0 approaches the average; this kind of value function is proven robust to a wider set of perturbations. It matches the widely used RL algorithm SAC in MuJoCo simulation and earns higher returns on real robots, staying steadier under perturbations to observations, environment and dynamics. The conclusion: regularization does narrow the sim-to-real gap.
4.40
Robot policies trained with reinforcement learning often jitter between actions, which wears out motors on real hardware. Existing methods directly limit how much actions change when the input state changes slightly, but these broad constraints hurt task performance as smoothing grows. Penalizing only the difference between consecutive actions was thought unable to handle observation noise. This paper proves that when several current states lead to the same next state, the consecutive-step penalty also shrinks the action differences between them. That amounts to smoothing across nearby states, and experiments show the policy becomes steadier under observation noise. CATS uses only the consecutive-step penalty and ramps its weight up linearly from zero, letting the policy learn to earn reward before it gradually smooths. In simulation and on real robots, action oscillation drops sharply with little loss in task performance.