DreamTrue’s robot world-model preprint is up. Mini packs are out; the GitHub license is not

What was posted on 8 and 9 October?

The arXiv Atom record we opened lists 2610.12468v1 as published at 17:59:51 UTC on 8 October 2026, primary class cs.RO, with a cs.CV secondary. The title is “DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training.” The HTML copy’s footnote says Xiangshuo Liu was an intern at CASIA during the work. The bibtex block on the GitHub README repeats the title, names the same nine authors, sets year 2026, and adds the note “Manuscript under review.”

The public GitHub repository brave-eai/DreamTrue did not exist on the paper’s posting day. GitHub’s repository JSON, retrieved 9 October, gives created_at 2026-10-09T02:35:37Z, pushed_at 2026-10-09T06:12:46Z, default branch main, language Python, stargazers_count 4, and license null. A GET of /brave-eai/DreamTrue/blob/main/LICENSE returned 404. The root README points to the paper, a project page at brave-eai.github.io/DreamTrue, and ModelScope dataset huoxingdawang/DreamTrue.

This is a video world-model paper, not a new robot product. For another 9 October latent-prediction write-up, see Meta FAIR’s RoboJEPA note. For a longer-horizon action-conditioned video model, see NVIDIA Long-WAM.

What problem do the authors say they are solving?

The abstract says existing robot datasets create two obstacles for a world model that must follow actions and keep contact physical. Imprecise camera calibration can throw rendered action conditions off the recorded video. Successful-only demonstrations can bias the generator toward completions even when the commanded motion should miss.

Their proposed fixes sit in two stages. Offline geometric calibration recovers camera parameters, and optional arm-mount offsets, from RGB demonstrations so rendered robot trajectories line up with the target frames. Counterfactual post-training then perturbs recorded end-effector goals, generates futures under those actions, and scores the videos with a learned embodied reward model instead of paired ground-truth futures.

They say the reward model is trained on human labels of robot, object, and interaction defects, and that they release 44.9 thousand annotated videos (30.4 thousand with defects) plus refined calibration for 153,666 episodes. Those counts are in the paper. We did not download the archive.

How do they say the generator is built?

Stage I converts each embodiment’s actions into a shared image-space bundle: rendered RGB, depth, amodal masks, gripper openings mapped to background intensity, and Plücker camera-ray maps. A VACE branch injects those features into a video DiT. The wmvideo README, opened the same day, names the public inference runtime as “a focused Wan2.1 V2V-VACE 14B runtime with a three-view launcher.”

Stage II builds counterfactual trajectories by applying an SE(3) perturbation to a recorded final pose, interpolating from the start, and solving inverse kinematics. A vision-language reward model, which the ModelScope card says is fine-tuned from a Qwen3.5 VLM, predicts defect probabilities on three axes. Those scores, plus a PSNR term on recorded actions, guide DiffusionNFT post-training. The paper’s own limitation paragraph says the action representation, generator, and VLM reward all operate in image space, so occlusion can still reward a visually tidy but physically wrong contact.

The public checkpoint is a dual-LoRA file, not a from-scratch 14B train. For what that adapter usually means, see LoRA research explained. For why a dataset license field is not the same as a complete source grant, see open weights versus open source.

Which numbers are author-reported?

Table 1 in the HTML is the authors’ AgiBot comparison against EnerVerse-AC, Genie Envisioner, DreamDojo, and GE-Sim 2.0, plus an ablation of their Stage I model without the reward stage. They report full-model nDTW of 0.8772 on recorded actions and 0.8831 on counterfactual actions, PSNR 23.06, SSIM 0.936, and LPIPS 0.094 on recorded video. They write that reward-guided post-training cut human-assessed object defects from 31.88% to 3.12% and interaction defects from 48.12% to 6.25%, with Fleiss’ κ 0.7817 on any-defect labels among three annotators.

Table 2 is their AgiBot World Challenge 2026 world-model-track scoreboard. They list themselves first on visual quality 0.6246, action following 0.9651, and overall EWMScore 0.829, ahead of teams they name Loop and PAIWorld. Scene consistency on that table is 0.8974, which is not the highest row they print. The abstract’s “ranks first” sentence is this table. We did not open an independent Challenge site.

On DROID they report PSNR 22.95 against a Ctrl-World reprint of 22.00. A RoboTwin policy-evaluation inset says success-rate mean absolute error fell from 12.92 to 7.08 percentage points. Those are author figures. Qualitative WidowX250 and Piper rollouts are described as zero-shot transfer. We did not watch those videos.

What can you actually download?

The root README splits the public tree into calibration, wmvideo, and reward. The wmvideo README says the dual-LoRA file lives at ModelScope path wmvideo/checkpoint-wmvideo.safetensors, that launchers run offline and do not fetch weights, and that Wan2.1-VACE-14B base weights must come from Wan-AI on Hugging Face or ModelScope. It names a condition mini package, condition/condition-mini.tar.gz, and lists default three-view layouts for AgiBotWorld-Beta, DROID, RoboMIND 2.0, and RoboTwin 2.0. Training code is not included in that component.

The ModelScope markdown we opened on 9 October, last_updated 2026-10-08, lists those two files plus reward/checkpoint/ and reward/reward-data-mini.tar.gz. A “Coming soon” block — echoed by the calibration README — says the full calibration set, a fine-tuned SAM3 checkpoint, and the full reward annotations are still pending. The same page prints license “Apache License 2.0” and downloads 25. A download counter is not a quality signal.

The project page at brave-eai.github.io/DreamTrue returned a JavaScript shell when we fetched it. We used the arXiv HTML, the GitHub README files, and the ModelScope markdown for the facts above. We did not clone the repository or load the LoRA.

What did we not verify?

This is an evidence review of the 8 October abstract, Atom record, and HTML, the 9 October GitHub API record and README files, and the 8 October ModelScope dataset markdown. We did not attend the AgiBot World Challenge, did not rerun nDTW or the human defect protocol, and did not confirm that every listed baseline used the same views, horizon, or prompt. Treat the method as what the paper describes, the scores as what the authors printed, and the download as a mini-pack inference drop plus a dataset license field that the GitHub repo does not repeat.