Robotics Briefing /
Learning a room, remembering a task, and choosing better experience.
Getting a robot to work reliably often means adapting to the place, remembering what has already happened, and learning from the right parts of its experience. This edition examines three approaches to those problems, with physical results that show both useful progress and narrow limits.
Your quick takeaways
- Splat2Act turns room scans into simulated pickup practice. Its strongest tested policy reached 63.9% success on a physical Stretch 3, but performance fell sharply in an unfamiliar room.
- Evo-RL combines human corrections with selected training experience. Its 88.24% insertion result measures individual attempts, not complete five-object runs.
- Dual-stream memory improves repeated actions more than object tracking. Average physical success rose to 62.5%, while performance on a cup-swapping task declined.
Papers
Splat2Act: a scanned room becomes a robot’s practice environment
Xiaomeng Ye and colleagues, Georgia Tech, Waymo and Toyota Research Institute • October 9 preprint • Catch-up
Rooms contain shelves, tables and objects that differ from a robot’s training examples. Splat2Act starts with a handheld scan and object images, reconstructs a visual and geometric copy of the environment, and generates pickup demonstrations in simulation. Those demonstrations adapt a vision-language-action model, which connects images and instructions to robot movements.
The pipeline processed 101 campus environments and produced more than 68,000 demonstrations. That is the scale of its simulated training collection, not the number of buildings tested with a physical robot.
On a Stretch 3 in the target office, the strongest personalized policy achieved 63.9% success across thirty-six physical attempts covering twelve tasks. A successful attempt required picking up the requested object and ending the policy’s execution. The same task subset reached 66.7% in simulation.
The limitation is how much this depends on the familiar scene. In an unseen lab, that policy reached 33.3% on one table and 22.2% on another. The unadapted model’s headline zero-success baseline also lacked exposure to the Stretch robot and its action format, so the comparison includes learning the robot itself. The work supports environment-specific adaptation; it does not establish general household fetching. Code and dataset links remain marked coming soon.
Evo-RL: keep the corrections and the useful parts of experience
Kai Ye and colleagues, Shanghai Jiao Tong University and Boden AI • October 9 preprint • Catch-up
A successful robot run can still contain hesitation, wasted motion and repeated attempts. Evo-RL combines human corrections with selected segments of demonstrations and successful autonomous runs. A learned value model estimates which actions improve the expected outcome, giving those segments more weight during training.
In physical screw-and-sleeve insertion, first-round Evo-RL achieved 88.24% attempt success, compared with 57.78% for RECAP. The difference is 30.46 percentage points. A second learning round reached 92.16%. The results count successful individual insertion attempts across a five-object sequence, rather than the proportion of complete sequences finished without retries.
The paper also compares how people spend collection time. With a four-hour active human-control budget, replacing one hour of demonstrations with targeted corrections raised attempt success from 27.03% to 51.52%. That suggests examples of recovery can be more informative than additional demonstrations of an already familiar route.
Evaluation covers insertion and shirt folding on two platforms. Human involvement remains part of the setup; folding errors were manually reset to an earlier stage. The active-control budget therefore should not be mistaken for the entire cost of supervision, setup and learning. Public code and dataset directories are available, although those resources predate the new paper.
Dual-stream memory: remembering progress is different from remembering an image
Yang Wang and colleagues, Magiclab Robotics and UCLA • October 9 preprint • Catch-up
The same camera view can require different actions depending on what happened earlier. A robot asked to repeat a movement twice needs to know whether it is on the first repetition or has finished. A robot searching for a hidden object needs access to an earlier visual clue.
This method separates those needs. One memory stream compresses execution history into a compact state; another retrieves relevant past images with more spatial detail. The two streams feed a manipulation policy together.
On four physical tasks using a Piper dual-arm robot, average success reached 62.5%, versus 37.5% for the FrameSamp+Modul comparison. Each method used the same four hundred demonstrations and thirty evaluation rollouts per task. The largest gains came from repeating picking and swinging actions the requested number of times.
The aggregate hides a setback: after three cup swaps, target selection fell from 53.3% to 40%. Retrieving an earlier image did not reliably preserve the target’s identity as the cups moved. Training checkpoints were also selected by similar loss rather than equal training steps. The results make memory’s role more specific: tracking task progress helped substantially, while persistent object tracking remains unresolved.
News
Danu’s recycling robot puts completed picks at the center of the sales pitch
October 9 reporting • Catch-up • Company product claims
TechCrunch’s corrected report says Danu Robotics is raising a $5 million late-seed round, rather than having closed it.
Danu’s own product description emphasizes fitting two-arm sorting robots into existing waste facilities. Its vision system identifies material on the conveyor and guides grippers toward the appropriate bin. The company says it records a pick only after the object drops into that bin.
That distinction matters: identifying an item or touching it is not the same as recovering recyclable material. Danu also argues for measuring successful picks over a working day, rather than short bursts of peak speed. These are sensible evaluation criteria, but the company page does not provide an independently verified, sustained-operation comparison. Retrofit claims and operating economics still need evidence from actual sites.
Read this first
Read the dual-stream memory paper’s physical-task table. The contrast between strong repetition gains and weaker cup tracking is a useful reminder that an improved average can hide an important capability getting worse.
Read the paper