Results at a glance
| What was measured | LiOps | Compared with |
|---|---|---|
| Object segmentation | ||
| Accuracy on indoor 3D scenes | 74.3 | 73.6PTv3 |
| Training rounds to peak accuracy (fewer is better) | 25 | 76PTv3 |
| Scan registration | ||
| Correct matches on scans with little overlap | 58.7% | 43.5%GeoTransformer |
| Correct among the 1,000 most confident matches | 88.7% | 78.6%SHReg |
| Site perception | ||
| Forklift pockets found | 99.4% | 78.1%Previous SOTA (internal architecture, unpublished) |
| Robot manipulation | ||
| Success on precise insertion | 86% | 0%Diffusion Policy · ACT |
| Success on cube stacking | 94% | 81%Diffusion Policy |
| Success on cube pushing | 100% | 89%ACT |
One AI for every 3D task
Until now each 3D task has needed a different AI. Even PTv3, regarded as the strongest in research, is weak at registering scans. LiOps leads on all three tasks with one AI.
More accurate than PTv3, and three times faster to learn
The task is to say whether each point in a 3D scene belongs to a chair or a desk. It works from shape alone, without colour, on a model 25% smaller than PTv3.
It registers accurately even with little overlap
The task is to join two 3D scans taken from different positions. That means pairing up the same points across both scans, and LiOps finds those pairs far more accurately.
Far ahead of a previous best that even the latest research SOTA (PTv3) could not beat
To lift a load, a forklift has to know exactly where the load and its fork pockets are. The previous SOTA is an internal architecture we built earlier (unpublished), which was more accurate than PTv3 at this job. LiOps is far more accurate still.
Previous SOTA: our earlier internal architecture, unpublished
LoadPrevious SOTA: our earlier internal architecture
Moving pointsBeyond 3D, to Physical AI
LLMs became general by learning how to answer questions rather than memorising fixed answers. Today's robot AI answers only “how should the joints move next?” LiOps answers “how will this scene move?”, and the robot's own motion sits inside that answer.
LLM
- Learns
- Question → answer
- Can answer
- Almost any question
- Scales
- Unchanged when the task changes
Existing robot AI
- Learns
- Image → joint motion
- Can answer
- The robot's own next move
- Scales
- Retrained for each robot and task
LiOps
- Learns
- 3D space + question → future motion
- Can answer
- How the robot and the things around it will move
- Scales
- The robot's motion is part of the world's motion
It predicts how a scene will move, to the millimetre
In scenes where objects move 56 mm on average, the prediction error is 3.8 mm. Learning this prediction alongside the task doubles the robot's success rate.
It does precise work other AI cannot
This is the standard test in robot arm simulation. It learned from demonstrations alone, with no pre-training. It succeeded 86% of the time on an insertion other AI never managed once, and leads dedicated models at cube stacking even after learning six tasks on one model.
How it was measured
- Object segmentation: ScanNet20 val, 312 scenes, coordinates (xyz) only, no pre-training, same Pointcept v1.7 recipe (100 epochs). Accuracy is best val mIoU; training rounds are the epoch at which that best was reached. Parameters: LiOps 34.8 M, PTv3 46.2 M.
- Scan registration: official 3DMatch (normal) / 3DLoMatch (low overlap) protocol, LiOps epoch 8. Correct matches is inlier ratio; the 1,000 most confident matches is IR@1000. GeoTransformer and SHReg figures are as reported in their papers.
- Site perception: one forklift validation scene (3-frame interval), LiOps single-frame final model (v10_single step 12,000). Perception accuracy is F1; pockets found is recall. The previous SOTA is our internal architecture before LiOps (unpublished, an A3DP model based on SPUNet-FlexGEMM). Its figures are those recorded in commit 1889e4eb and were not run side by side on the same frames. The previous SOTA figures are after threshold tuning; the LiOps figures are before. The comparison of the previous SOTA with PTv3 is an experiment in PointWorld swapping only the backbone (DROID, 3,185 clips, same head, loss and evaluator).
- Motion prediction: 300 ManiSkill validation clips. The error of assuming no motion equals the average size of motion in the scene. The learning-alongside comparison uses the same model and 30k steps on PegInsertionSide.
- Robot manipulation: ManiSkill3, 1,000 motion-planning demonstrations, 100 episodes, counted as success if it succeeded at least once in the episode. LiOps takes RGB-D point cloud input. Precise insertion (PegInsertionSide) uses the Peg/Plug model; cube stacking (StackCube) and pushing (PushCube) use a single model shared across six tasks. Diffusion Policy and ACT are the RGB baselines in Table III of the ManiSkill3 paper (best success during training). Diffusion Policy trained directly by us on the same evaluator reached 2% on precise insertion.