Robotics Briefing /
Knowing when to yield, and what to learn next.
A robot carrying a box with a person needs to give way without losing its grip. A robot learning a complex job needs to practice the steps that actually fail. Today’s research examines both problems, while new industry demonstrations shift attention toward completing whole workflows.
Your quick takeaways
- CEER2 lets a humanoid yield in one direction while remaining firm in another. Physical demonstrations include writing and carrying a three-kilogram load with a person; reliable contact safety remains unproven.
- RoboCoach uses predicted failures to request targeted human demonstrations. With 150 additional demonstrations per platform, physical task success reached 75% on Franka and 83.8% on AgileX.
- XPENG’s IronMind reached 55% success across six unfamiliar manipulation tasks after large-scale pretraining and robot-specific training. Adding world-model supervision reduced physical success to 31.7%, despite improving a separate prediction test.
Papers
CEER2: yielding to contact without giving up control
Xinyuan Luo and colleagues, Duke University • September 30 • Preprint with physical demonstrations
Making every part of a robot equally soft can create a new problem: the object it is holding slips. CEER2 controls how readily the hands yield along different directions, while separately selecting whether the body resists a push or follows it. A learned controller modifies an existing whole-body motion tracker and estimates interaction forces from the robot’s own movement history.
In physical tests, the system carried a three-kilogram load with a person by combining a yielding contact direction with firmer support directions. An equally soft setting lost the box. During writing, directional compliance maintained useful contact; a uniformly stiff comparison broke the pen holder in another trial.
The practical point is that handling contact requires choosing where to yield, not simply reducing force everywhere. The system also changed hand stiffness while operating, without restarting its underlying controller.
These are demonstrations rather than a large reliability study. Manual force measurements showed substantial differences between requested and measured stiffness. The authors also report heavy footsteps and insufficiently smooth physical interaction. Those limits matter for working around people, and the results do not establish general safety. Project code is marked coming soon.
RoboCoach: using predicted failures to choose better practice
Jiajun Liu and colleagues, Tsinghua and collaborators • September 30 • Preprint with physical tests
RoboCoach asks which part of a multi-step job needs more teaching. Its world model predicts how an attempted action sequence will unfold. A progress checker identifies the first unfinished step, then the system requests human demonstrations for that skill and updates its corresponding controller component.
After 150 additional demonstrations per platform, complete-task success rose from 13.3% to 75% on Franka and from 40% to 83.8% on AgileX. A comparison that spread demonstrations uniformly and updated a shared controller reached 30% and 47.5%. Evaluation used twenty trials per task across three Franka tasks and four AgileX tasks.
The results suggest that both selecting useful practice and directing it to the relevant skill can matter. The physical comparison changes those two choices together; separate experiments isolating their contributions were conducted in simulation.
Transfer remained harder: four previously unseen combinations of skills averaged 35% success across twenty trials each. Predictions can miss hidden contact or events outside the camera’s view, and the skill library is predefined. The system still requires new human demonstrations. Public CoachWorld code and weight files are available for the predictive model.
IronMind: larger human-video training helps, but prediction is not enough
XPENG Robotics • September 30 • Technical report with physical tests
Human hands and robot hands move differently. IronMind represents their actions relative to the camera, reducing the need to reconstruct a human’s full body and map it directly onto the robot. Its training combines carefully filtered first-person human video with data from other robots.
After robot-specific training, the model pretrained on 10,000 hours reached 55% success across six unfamiliar manipulation tasks, with ten trials per task. These included new objects, grasping requirements and spatial instructions. Models pretrained on up to 5,000 hours reached at most 11.7% under the same study.
An especially useful comparison holds the data budget fixed: expressing actions relative to the robot’s torso instead of the camera reduced success to 26.7%. How the data describes movement matters alongside its quantity.
Another finding complicates the story. Additional supervision from world-model and visual prediction objectives improved a separate trajectory-prediction evaluation, but reduced physical task success from 55% to 31.7%. Better predictions did not automatically produce better actions.
This is evidence from one robot embodiment, six tasks and small trial counts. Camera calibration and training on the target robot remain necessary. The project provides demonstrations; no public code or model-weight release was verified.
News
Destro raises funding for coordinating mixed warehouse fleets
September 29 announcement • Industry catch-up • Company-reported deployment
Destro announced an $8 million seed round for software coordinating robots, workers and warehouse systems. MothershipOS assigns and sequences work across the operation; VisionOS handles perception and manipulation on the robots.
The company says it is operating in production with logistics providers, and its release includes a supporting statement from Yusen Logistics. The potential value is fewer gaps between individual automated steps, especially when equipment comes from different manufacturers.
The release does not provide independently verified throughput, uptime or intervention rates. Funding and a customer statement establish commercial activity, but do not quantify how reliably the system handles a full warehouse shift.
Highlights
Dyna shows an hour-long laundry workflow
September 29 • Demonstration catch-up • Company-reported autonomy
Dyna introduced its wheeled Taku robot and Dyna-2.1 system with footage it describes as an uncut, autonomous hour of laundry-room work. A high-level decision system selects steps, while the action model and whole-body controller coordinate movement and handling.
The relevant advance is the scope of the demonstration: completing connected steps without someone continually refilling or clearing a single workstation. That is closer to the operational burden customers care about than an isolated successful grasp.
The technical report identifies customer-site rollout of this new system as its next milestone. The demonstration therefore should not be mistaken for independently validated, repeated production shifts. Published failure distributions and sustained site-level results would make the claim easier to assess.
Read this first
Read IronMind’s physical evaluation and Table 5. The drop in task success after adding prediction supervision is a clear example of why robot progress must be checked on the physical task itself.
Read the paper