A robot policy called GaussianDream++ reported higher success than a reproduced π0.5 policy across the main benchmark comparison and in a pooled physical-robot evaluation. The reported advantage came with a measurable cost: each action chunk took 330 milliseconds for GaussianDream++ compared with 286 milliseconds for reproduced π0.5, a difference of 44 milliseconds.
On the standard LIBERO benchmark, GaussianDream++ reported 98.6% average success. On LIBERO-Plus, which measures performance under controlled changes to the task setting, its reported Overall score was 87.8%. In the physical evaluation, pooled success was 52.5% for GaussianDream++ and 29.2% for reproduced π0.5, an absolute gap of 23.3 percentage points.
Tokens carry a compressed picture of the scene
The system places two types of learned tokens directly in the PaliGemma backbone. World State Tokens represent the current scene, while World Prediction Tokens represent future states. Both share the visual-language context and remain visible to the Action Expert.
The main configuration uses 16 World State Tokens and four World Prediction Tokens. They act as compact internal summaries rather than one-for-one replacements for the Gaussian primitives used in the training representation.
The extra world representation is used during training. A World Representation Head decodes a Current World and a horizon-dependent Future Prediction over shared Gaussian primitives, using a static and dynamic factorization. At deployment, the head, renderer and auxiliary branches are removed, leaving 20 world tokens. The deployed policy does not perform online Gaussian decoding, rendering or future rollout.
The advantage was uneven across controlled shifts
The benchmark evaluation covered four standard LIBERO task suites and seven controlled LIBERO-Plus shifts. Policies were trained on standard LIBERO and evaluated on LIBERO-Plus without robustness-specific adaptation.
In the matched GaussianDream-family comparison, GaussianDream++ was 0.2 percentage points above GaussianDream and 1.7 points above reproduced π0.5 on LIBERO. On LIBERO-Plus Overall, it was 0.8 points above GaussianDream and 2.3 points above reproduced π0.5.
The result varied by the type of shift. Under Camera, GaussianDream++ reported 80.1% success versus 73.2% for reproduced π0.5. Under Layout, the figures were 90.0% and 87.7%, and under Noise they were 94.2% and 89.9%. GaussianDream++ was 2.8 points above GaussianDream on Camera and 1.6 points above it on Layout, while its performance under Robot shift remained below reproduced π0.5.
The paper limits direct numerical comparisons to the matched GaussianDream-family block. Published baseline results may differ in their training data, checkpoint selection and implementation, making broader comparisons less direct.
Physical results came from a small set of trials
The physical evaluation used a dual-arm robot, two manipulation tasks and three conditions, with 20 trials for each task-condition pair. Across the reported trials, GaussianDream++ reached 52.5% pooled success, compared with 29.2% for reproduced π0.5.
The reported figures were higher for GaussianDream++ in both tasks and in the Layout and Camera comparisons described in the analysis. For Bowl-Proximity, pooled success was 46.7% for GaussianDream++ and 25.0% for reproduced π0.5. For Eggplant-to-Pink-Plate, it was 58.3% and 33.3%. Under Layout, the figures were 50.0% and 25.0%, while under Camera they were 42.5% and 22.5%.
The supplied analysis reports no confidence intervals or inferential tests for the pooled result, and no subgroup uncertainty estimates for the task and condition comparisons.
Component tests traced a stepwise pattern
The ablation variants kept the PaliGemma and π0.5 backbone, action objective, training budget and evaluation protocol the same, changing only the specified component.
In the reported structural sequence, LIBERO-Plus Overall was 85.5% for the baseline, 86.3% after adding World Tokens without Gaussian supervision, 86.9% with Current World only, 87.2% with Current and Future Prediction uncoupled, 87.5% after coupling without static consistency, and 87.8% for the full model.
A separate supervision comparison reported Overall scores of 87.8% with metric depth and 86.9% without it. Camera success was 80.1% with metric depth and 76.5% without it, while Layout success was 90.0% and 87.7%. Removing metric 3D flow corresponded to an Overall score of 87.2%.
The comparison leaves room for further testing
The latency comparison was measured in a matched local pipeline: 330 milliseconds per action chunk for GaussianDream++ versus 286 milliseconds for reproduced π0.5. GaussianDream++ therefore took 44 milliseconds longer, or approximately 1.15 times the reproduced π0.5 latency. The GaussianDream reference was measured in a potentially different environment, so the supplied analysis does not establish an exact speedup over that system.
The reported evaluations covered two physical tasks, three conditions and 20 trials per task-condition pair, along with four LIBERO suites and seven LIBERO-Plus shifts.
The work is an arXiv preprint, version 1, dated 26 Aug 2026. Funding information is not reported in the supplied document.
Paper data and sources
Original title: GaussianDream++: Efficient 3D Gaussian World Modeling for Robotic Manipulation
Authors: Yuqing Jiang, Zijian Zhang, Weitao Zhou et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text