EMBODIED AI · 2026-11-17
Embodied AI — Paper Daily, 2026-11-17
Today's embodied track asks how a policy and its data can move more reliably from simulation or one robot to hardware and to another robot. One paper uses reinforcement learning in simulation only, so an open-weight reasoning model learns to output end-effector actions and then moves to a real dual-arm robot. Another takes a bipedal humanoid through a long indoor-to-outdoor mission, first adapting the model in a site twin. A locomotion recipe is written in a shared, body-aligned joint space so it can be reused across humanoids, and object navigation revises map errors from predicted depth with new views. Other work grows further-training demonstrations from a model's failures, or treats visual failure as a switch back to proprioception. On contact-rich manipulation, one paper has a language model write an objective program a planner can revise, another uses tactile recordings to recover force that hand pose cannot show, and a third finds that no single pretrained visual representation is best everywhere on insertion.
4.50
Open-weight reasoning models asked to control a real robot usually still lag far behind frontier proprietary models, and they often need demonstrations or an external harness. The authors use reinforcement learning in simulation only, so a 27B open-weight reasoning model observes the environment, writes a chain of thought, and outputs the next end-effector pose, with no real-robot demonstrations. In simulation, success on LIBERO rises from 1.6% to 74.4%, approaching frontier proprietary models at high reasoning effort. On a real dual-arm robot it succeeds in 60% of trials, while the untrained base model succeeds in none, and it is more robust to a change of task than an action model fine-tuned on demonstrations. A model can first learn to use a robot in simulation, then move to hardware, without stockpiling demonstrations first. For general-purpose robot agents, robot use can be part of foundation-model post-training.
4.49
AAGIBOT
A bipedal humanoid walking from indoors to outdoors on a long language-instructed mission passes rooms, doors, and roads, and the camera moves with the gait; few systems have actually finished such a route. The authors navigate with only the built-in color camera: one model predicts actions, another predicts a goal, and an agent workflow splits a short instruction into sub-tasks and adds obstacle avoidance. On the Lingxi X2, it completes a mission of about a hundred metres from an office lobby to a store next door, can change destination, and goes around obstacles. After adaptation on data from a digital twin of the site, closed-loop success in the twin rises from 40% to 80%, and the model is then deployed on the real robot. For a long indoor-to-outdoor mission, first split the instruction and adapt the model in a twin of the site. For teams that want a humanoid to walk itself to a workplace, this is a deployable pipeline, not only a higher score on simulated benchmarks.
4.47
Humanoid walking policies often output native joint angles, and rewards and action limits are written in those robot-specific coordinates, so a recipe tuned on one robot is hard to reuse on another. The authors add a body-aligned virtual joint space in front of the existing low-level control: the policy follows a shared anatomical convention, then a map sends the actions to each robot's native joints. With a single unmodified recipe on three humanoids whose leg-axis layouts differ, only this interface tracks velocity accurately, keeps the feet aligned with travel, and trains stably on all three. On a Unitree G1, heading yaw rate falls from 5.31 degrees per second in native coordinates to 0.89, about one sixth of the original drift. Actions have to be written in a shared, body-aligned coordinate system before a training recipe can be reused across humanoids. For teams reusing locomotion training on several humanoids, this is less work than retuning rewards on every robot.
4.45
Depth predicted from color images can support object navigation without a depth camera. If that prediction is written into the map as if it came from a sensor, view-dependent errors accumulate, and the search ends before new views can correct them. The authors build a system that needs no further training: the robot moves to revise the map with new views, still searches from reachable viewpoints when no frontier remains, and stops only after confirming the target across views. On three benchmarks, it is the strongest method that uses no depth sensor and no extra training. Plugging the same predicted depth into an existing system raises success from 10.8% to 61.7%, and the system was also deployed on a real robot. Errors in predicted depth have to be corrected by new views from movement, not written into the map once as reliable readings. For indoor navigation with a lighter sensor suite, the map must stay revisable and exploration must be able to continue.
4.15
Improving a pretrained vision-language-action model requires demonstrations that address its failures, not just any successful run. Coding agents can roll out trajectories on their own, but finishing a task does not mean the data are fit for further training. The authors build a failure-driven data engine: it analyzes failed rollouts, selects intervention states, iteratively writes closed-loop code policies, then collects successful continuations from states where the model struggles, adding noise, dropping idle frames, and filtering with the base model's loss. In simulation, success on LIBERO-10 rises from 79.4% to 87.8%, the other two suites also improve, and the updated policy no longer needs the coding agent. Demonstrations for further training should grow from where the model fails, and should be filtered first. For post-training such a model, this ties finding weaknesses to creating and curating the data in one pipeline.
4.14
Imitation policies are often trained on noisy object-position readings, and they learn to discount those readings. Because demonstrators steer toward the object, the robot's own state already carries information about where the object is. A trained policy combines this proprioceptive anchor with the reading, and gives the reading less weight as the noise grows. The anchor's error belongs to the learner and is much smaller than a linear estimate from the demonstrations, so the classical method overestimates that weight. Using the measured anchor error to predict sensitivity in held-out conditions has a mean absolute error of only 0.025, against 0.37 for the classical method. The discount does not carry over as is: in closed loop the policy produces its own state, and it is more sensitive to a persistent reading offset. All of these results are in simulation, with synthetic noise. The discount applied to noisy readings in training should not be assumed to hold in the same way at deployment. For imitation learning under noisy perception, closed-loop sensitivity to bias needs its own check.
4.12
Legged robots that use vision for perceptive locomotion can move quickly over complex terrain, but when occlusion, reflections, or a sudden lighting change make vision fail, the motion often collapses with it. The authors treat severe perception failure not as ordinary noise, but as a condition for switching off exteroception and falling back on proprioception. A teacher learns both vision-driven motion and a proprioceptive fallback, and a student estimates exteroceptive features and their uncertainty, then suppresses them on out-of-distribution inputs. On a Unitree Go2 with a depth camera, the robot can cross large gaps, raised platforms, and narrow beams when vision is normal, and it still walks stably under heavy occlusion and strong lighting changes. When vision is unreliable, exteroceptive features should be switched off and the robot should fall back on proprioception, rather than keep trusting a broken observation. For quadruped locomotion outdoors or under difficult lighting, perception failure becomes a switchable control mode.
4.09
Connecting a vision-language model to model-predictive control can combine semantic reasoning with physics-based planning, but contact-rich manipulation still needs a task representation that both sides can read. The authors have a pretrained vision-language model write an editable objective program. Stages specify costs and completion conditions, a planner executes and revises the program in a digital twin rebuilt from the real scene, and it is then deployed on the robot, without updating model weights. In simulation on four tabletop tasks, success reaches 90% after two feedback rounds. Programs that succeed in simulation, run on the robot with object locations changed, succeed in 32 of 40 hardware executions. A language model need not output torques directly; it can first write an objective program that a planner can execute and revise. For contact-rich tabletop manipulation, this lands a semantic instruction on a cost program that can be tried in a twin and then moved to the robot.
3.87
Visuomotor imitation for real-robot manipulation is often data-hungry and easily thrown off by visual changes that do not matter to the task. Pretrained visual representations are supposed to help, but which is best for robots is still unsettled, and existing conclusions often conflict. The authors narrow the question to contact-rich insertion, compare widely used vision foundation models with an explicit contact-centric representation, perturb each along occlusion and tolerance, and test transfer to a real robot. Even between two nearly identical insertion tasks, the best representation changes, and small changes make the trend hard to predict. Even so, both simulation and hardware show that spatially grounded, especially contact-centric, information is the most stable help for such policies, though such cues are hard to estimate reliably under real visual complexity and sensor noise. Contact-rich manipulation should not expect one pretrained visual representation to be best everywhere. For insertion and assembly, it is more worth investing in perception that can estimate contact reliably than in swapping in a larger vision model alone.
3.86
Dexterous manipulation relies more and more on human demonstrations, which a simulation pipeline then multiplies. Such pipelines usually rebuild from vision how the hand moves, but which fingers touch and how hard remain unobserved, because a hand pose shows where a finger stops, not how far it was driven to close. On held-out recordings, a policy that only imitates the observed pose grips with 0.31 times the recorded force. The authors record real contact with a tactile glove, replay the contact and closing force in simulation, and a policy trained on those replays grips at 0.99 times the recorded force, with more accurate contact prediction as well. Where the hand stops cannot recover how hard it pressed; tactile recordings are what close that force gap. For teams that want to multiply dexterous demonstrations, a simulated replay should match contact and force, not only the visible hand pose.