Robotics Briefing /
Practice in simulation, proof on real objects
Changing the shape of a mug or the light on a table can undo hours of robot practice. Two research groups tested different ways to make simulated practice carry over to real objects. A third team measured what a hand gains when it can feel contact. On a factory floor, Google DeepMind and Apptronik put a humanoid to work sorting metal brackets.
Your quick takeaways
- Starting with one simulated demonstration per task, researchers generated enough practice for a two-armed robot to score 84.5% to 90% on three physical tasks. Training on 60 real demonstrations per task yielded scores near 50%. Two tasks gave partial credit.
- Google DeepMind says Gemini Robotics 2.5 helped a humanoid pick and pack car parts with 96% success in a proof-of-concept on Mercedes-Benz production lines. The post does not say how many attempts or how many hours stand behind that figure.
- Across 26 physical tasks, adding fingertip touch features improved four robot controllers by about 14 to 25 percentage points. One went from 48.3% success to 73.1%. The team lists two model checkpoints and has not released the training data.
Papers
One demonstration, thousands of mugs
Tianyuan Dai, Kevin Lin and colleagues at UT Austin and NVIDIA · November 8 paper listed on alphaXiv · More than 3,200 physical trials
Robots trained in simulation tend to fail on real objects whose shapes they never practiced on. SimScale asks software agents to write a small program for a whole category, such as mugs, after searching the internet for examples. The program has dials for body height, handle shape and so on, and each setting produces a new mug. Two reviewing agents reject versions that miss common shapes or produce broken ones, such as a handle that floats. A second set of agents then adapts a single human demonstration, recorded in simulation, to each new shape. That yields 4,000 practice runs per task.
Policies trained only on that data were tested on a physical two-armed robot with five test sets of objects per task and 20 attempts per set. With one of the two policy types, SimScale scored 90% on placing a pot, 84.5% on hanging a mug and 84.5% on putting a charger in a drawer. The same policy trained on 60 real demonstrations per task scored about 50% on each. One more result deserves attention. Programs written directly by a language model solved the tasks in simulation more than 95% of the time and almost never on the real robot. They moved at odd speeds and acted when the wrist camera could not see the target.
This is a short paper with three tasks, all using rigid household objects. Two of the tasks award partial credit for steps along the way, so a score is not a plain success rate. The real-demonstration baseline saw only three of the five object sets. The project page offers result files but no code.
Teaching a robot’s eyes to ignore the lighting
Sunshine Jiang and colleagues at Amazon FAR, Stanford, UC Berkeley and Georgia Tech · November 9 preprint · Simulation and physical tests
A simulated kitchen never looks quite like a real one, and a policy that learned from simulated images can be thrown by real light and real surfaces. The usual fix is to scramble colors, textures and lighting during training so the real world looks like one more variation. Every new task has to repeat that work. INVAR does it once. A simulator can render the exact same moment twice, once clean and once scrambled. The researchers trained a vision model on about 6.3 million such pairs so that both versions produce the same internal description. Later policies keep that vision model fixed and train on clean simulated images only.
On a physical robot hand with 22 joints, policies built this way succeeded 61% of the time across grasping five unfamiliar objects and pouring into a bowl. The strongest alternative, trained with scrambling, managed 30%. On a two-armed robot with simple grippers the figures were 65% and 32%. Grasping results rest on 25 attempts per object and the other tasks on 24 attempts per method.
The method handles appearance only. The authors say differences in shape, physics and motor behavior between simulation and reality remain unaddressed. The project page does not link code or model files.
One sense of touch for many kinds of fingertip
Bing Han and colleagues at Robbyant · November 9 preprint · Simulation and physical tests
Robot fingertip sensors work in different ways. Some measure electrical resistance, some magnetic fields, and some point a tiny camera at a soft gel. A model trained on one type cannot read another, so touch data has stayed in small, separate piles. LingBot-Touch maps readings from all three types onto a common fingertip layout and trains one small model on 5,000 hours of recordings from sensor gloves worn by people and from robot hands steered by operators.
The model’s output can be added to an existing policy without retraining the model itself. Across 26 physical tasks chosen because touch matters, such as ejecting a memory card or moving an egg, it raised average success for four different policies by about 14 to 25 percentage points. For one of them the average went from 48.3% to 73.1%. Feeding the same policies raw sensor readings gave at most about half that gain. Each task got 20 attempts.
The tests used two hand designs. The model covers fingertips only, so contact on the palm or along the fingers is poorly represented. The public code computes touch features from sensor readings, and two checkpoints are listed on Hugging Face. The code page says access to the model files requires permission. No training data has been released.
News
Gemini Robotics 2.5 arrives with a factory trial and few details
Google DeepMind · November 10 announcement · Company-reported physical tests
Google DeepMind calls Gemini Robotics 2.5 its most capable model for turning camera images and instructions into robot movements. The company says it tested the model on 200 tasks and saw success rates 30% to 100% higher than earlier baselines, depending on the type of task, with tasks finished 20% to 30% faster. Those are improvements relative to older models, not success rates.
The post also describes a proof-of-concept on Mercedes-Benz production lines in Germany and Hungary. There, Apptronik’s Apollo 2 humanoid sorted car parts from bulk bins into trays with a reported 96% success rate.
DeepMind ran or reported all of these tests itself. The post gives no number of attempts and no duration for the factory figure, and it links no technical report. The model is rolling out to the company’s robotics partners. No public release is described.
Boston Dynamics puts Atlas on scooter wheels to test its motors
Boston Dynamics · November 9 engineering report · Physical research prototype
Boston Dynamics engineers swapped the feet of an Atlas humanoid for off-the-shelf scooter wheels, each mounted on the existing ankle motor with one custom adapter. The company is clear that this is a research rig and no product. Rolling pushes the ankle motors through their full range of speed and force on every launch, turn and stop. That made it a hard test of whether a controller learned by trial and error in simulation behaves the same on real motors.
The post says a learned controller can pick up moves the real hardware cannot perform, so the team tuned the simulated motor to match measurements from the robot. It reports that a 360-degree jump then worked on the first try. The post gives no measurements and describes no failures, so this is an engineer’s account and not a study.
Highlights
A robot model trained without robots, open to requests
AIRoA · Catch-up: November 6 release, technical report first submitted November 2
AIRoA-0 learned from 16,151 hours of people doing tasks with a handheld gripper. No robot was steered to collect its training data. The group reports 78% average success across ten two-armed tasks, 15 attempts each, against 27% for the best of four other models trained on the same data. The model weights are on Hugging Face behind an access form, under Google’s Gemma terms of use. The training code is public. The dataset is promised with no date.
Read this first
How can you tell whether a robot remembers how many times it picked up a cube? In a November 10 preprint, Chubin Zhang and Bo An interrupted a simulated robot twice for five seconds. A controller that normally finished the counting task in 47 of 50 trials now managed 16. When the robot still had to wait but the pause frames were kept out of its memory, it finished 45. Figure 1 shows the arm pressing the “done” button a whole repetition early. The experiment isolates a failure in how this controller stores its history. It has yet to be tested on physical hardware.
Read the paper