EMBODIED AI · 2026-11-05
Embodied AI — Paper Daily, 2026-11-05
Today's embodied track is mostly about where robot data comes from. AIRoA-0 trains a bimanual manipulation model only on demonstrations people recorded with a handheld gripper, the soft-finger gripper paper shows how to collect such data better, GUILD asks which data to collect, and Liber-0 learns straight from human video and glove data. The other thread is making large-model abilities affordable on robots. ARP lets the large model update slowly while a small model outputs actions fast, AFO distills visual foundation models into a small RGB-only policy, and ZAP uses simulation to teach a small VLM the motion control it lacks.
4.79
Robot imitation learning usually starts with teleoperating a robot to collect data; AIRoA-0 collects no robot data at all. It is trained only on 16,151 hours of bimanual demonstrations that people recorded with a handheld gripper, and the robot shows up only at test time. At a site where none of its training data was recorded, using just one camera on each wrist, it sorts small parts, swaps the mop pads and brush of a robot vacuum, folds a towel and more, averaging 78% success over ten tasks. Post-trained on the same task demonstrations, the best of four other vision-language-action models (VLAs) reaches only 27%. Data collection no longer has to be tied to one robot: any robot that mounts the same gripper can run the model.
4.72
Fine-tuning a generalist vision-language-action model (VLA) on a few demonstrations of one task makes it better at that task, but it then fails more easily when the object, background or wording changes. RAFT splits post-training into two steps: it first trains a set of LoRA adapters on data annotated with task-relevant reasoning, such as which objects and regions matter, then freezes them and fine-tunes the policy on action demonstrations. On a physical DROID robot, success in situations unseen during training rises from 27.9% with plain fine-tuning to 75.0%. The resulting policy also attends more to task-relevant regions: teaching it where to look before teaching actions loses much less generalization.
4.50
Frontier models such as GPT-6 Astra can already control robots directly and generalize better than VLAs. The authors argue this ability splits into two skills: understanding what a task asks for, and actually moving the robot. Small open VLMs already have the first and mainly lack the second. ZAP teaches that missing skill entirely in simulation, on simple primitive tasks, using the model's own successful attempts, with no real-robot data and no distillation from a frontier model. On real-world tasks it reaches 96.7% success, matching GPT-6 Astra, while each task costs one seventeenth as much to run. By training only the missing skill, a small model gets close to a frontier one.
4.50
As VLAs grow, every action waits for the large model to finish, so the robot reacts slowly. ARP runs the two jobs separately at their own pace: the large VLM refreshes its understanding of the task and scene every so often, which the paper calls semantic memory, while a small vision-action model reads that memory and outputs actions at a high rate. Because the large model's view always lags a little, the small model is told at every step how old the memory is and how long its own actions will be delayed. During training it is also deliberately fed older memories, so it learns to act well despite the mismatch. On a Franka arm, compared with conventional VLA baselines, it succeeds more often and finishes tasks in 20% less time. The large model need not be present at every step: running the VLM at a tenth of its usual frequency costs almost nothing.
4.49
Dexterous manipulation is not only about moving an object somewhere; the fingers must keep holding it in a useful way so the next action can follow. Many methods specify only the object's target pose and leave the grasp to the policy. DexHOT writes the goal directly as the hand's joint posture plus the object's position relative to the wrist, so the policy learns to move from one grasp configuration to the next, and long tasks become a chain of such subgoals. All training data is synthetic, with no human reference motions. Across 120 objects, a single policy succeeds 80.5% of the time on seen objects and 69.4% on unseen ones.
4.49
When a user asks a mobile robot to put the toys away in the toy box, it must first work out which objects are toys and which box the user means. If the scene representation was built before the instruction arrived, it may simply lack the information needed to answer that. FEIJOA lets a language-model agent look things up after it gets the instruction: it repeatedly queries a grounding tool over the scanned scene, checks the returned candidates and rewrites its query until the target is clear. On open-ended grounding queries it gets 94.2% exactly right, against 68.0% for the best prebuilt representation. In real tasks on a Spot robot, most failures come from physical skills such as grasping, not from misunderstanding the instruction.
4.47
Visual foundation models for depth or segmentation give robot manipulation useful information, but they are too slow for closed-loop control and their outputs flicker from frame to frame. AFO uses them only during training. It first trains a teacher policy that sees both the RGB image and the outputs of several foundation models, then distills what the teacher learned into a student policy that sees only RGB. Only the student runs at deployment. On simulation benchmarks the student has over 30 times fewer parameters than foundation-model encoders and 50 times lower latency than its teacher, while still beating earlier foundation-model-based methods. It is a cheap way to get their perception without paying their latency.
4.47
Handheld grippers make grasping data fast and cheap to collect, but common designs either need external base stations or have rigid fingers that show nothing about grip force. This work builds a battery-powered handheld gripper on YUBI whose soft fingers wrap around objects and visibly bend under load, so the camera image shows how hard it is squeezing, and whose finger shape matches the robot's gripper. The authors use it to collect a multi-object pick-and-place dataset and train π0.5, with two practical findings. A top camera raises success by 13 points even when only about half the data includes it. Mid-training on this multi-task data before fine-tuning on new objects raises success from 35% to 57%.
4.17
Imitation-learning data for robots is usually collected by teleoperators going on intuition, and much of it does little for the policy. Earlier work filters the data afterwards, but the useless data has already been paid for. GUILD steers collection as it goes. A VLM first proposes which environments, object layouts and goals to record for the deployment setting. After collection, it trains a policy once on the data and uses data attribution to estimate each trajectory's effect on that setting, that is, whether adding it to training raises or lowers performance. That estimate then decides what to collect next. In simulation and on real robots it is more data-efficient than passive collection, several baselines and experienced roboticists. A version that only uses the VLM's proposals without this feedback averages about 25% lower: what to collect should depend on its real effect on the policy.