Three preprints released from 20–26 October 2026 shift motion generation away from producing an attractive sequence in isolation. SMART makes a selected part of a movement editable through conversation; Joints, Camera, Action! generates rig motion together with the video and camera that place it in a scene; ViT-H uses a robot's tracking outcome to improve the motion generator itself. All are Tier 3 preprints, so the claims below remain the authors' reported results rather than independently established findings.

SMART edits a phrase without discarding the rest

Text-to-motion systems usually answer a prompt in one pass. If the left arm is wrong, a user often has to regenerate the whole sequence and risk losing the parts that worked. Learn to Modify, submitted 20 October, proposes SMART, a role-specialised framework that interprets feedback, identifies the affected joints and time span, edits that masked region and evaluates the result. A memory component retains successful editing strategies across interactions.

The important change is the unit of control. Instead of treating a motion as one indivisible output, SMART treats it as a revisable composition with local continuity obligations. That makes the work relevant to choreographers and animators, but the paper's abstract does not establish how the system handles ambiguous embodied instructions such as “let the reach arrive with more hesitation.” The distinction echoes the Lab's account of why movement language and model coordinates do not line up automatically.

Source tier: Tier 3 — arXiv preprint.

Joints, Camera, Action! co-generates what moves and how it is seen

Joints, Camera, Action!, submitted 20 October, starts from a practical gap in video generation: a model may depict a convincing jump, yet not return a replayable 3D animation for a supplied asset. The system keeps an arbitrary-rig motion model separate from a pretrained video model and joins them with cross-attention. It also learns joint locations in the frames so that it can recover the camera and register the generated motion to the scene.

The authors report training with nearly 200,000 pairs assembled from edited 3D assets plus human and robot video. The separation matters because motion-only data is easier to obtain and augment than paired video and rig animation. For creative systems, the larger implication is that viewpoint is becoming part of the generative contract rather than an after-the-fact rendering choice. This extends the archive's earlier discussion of camera-centred motion representations from observation to co-generation.

Source tier: Tier 3 — arXiv preprint.

ViT-H lets execution failure teach the generator

ViT-H, submitted 23 October, converts monocular human video directly into humanoid reference motion with a diffusion model. It bypasses the usual separately optimised reconstruction, retargeting and tracking stages. More unusually, it uses outcomes from the downstream motion tracker as preference feedback to post-train the generator.

The authors report better motion quality and executability than the evaluated multi-stage pipelines on a Unitree G1, and describe an autoregressive variant for video teleoperation with latency below 400 milliseconds. Those results need replication, and “real time” here should not be confused with imperceptible latency in a performance loop. Still, ViT-H makes a consequential design move: a beautiful reference that a body cannot enact is treated as a generation error, not merely a controller problem. It complements the Lab's earlier account of why physical limits belong inside motion evaluation.

Source tier: Tier 3 — arXiv preprint.

The week's pattern: generation is becoming accountable downstream

Across editing, animation and humanoid control, the output is no longer judged only at the model boundary. It must survive a local revision, align with a camera and scene, or execute on a physical platform. This is a useful direction, but it also widens the evaluation burden: continuity, registration and feasibility now need to be tested alongside visual plausibility.

References

  • He, Z., Bian, T., Yang, Z., Li, H., Jiang, N., Cui, J., Jia, B., Wang, Y., Huang, S., & Zhu, Y. (2026). ViT-H: Generative video-to-humanoid control with motion tracking feedback. arXiv. https://arxiv.org/abs/2610.32966
  • Li, J., Zhao, J., Zhou, Y., Zhao, Z., Wu, S., Wang, Y., Wang, D., Chu, H., & Li, F. (2026). Learn to modify: Self-evolving multi-agent refinement framework for conversational text-to-motion generation. arXiv. https://arxiv.org/abs/2610.28210
  • Rotstein, N., Tewel, Y., Tessler, C., Maron, H., Chechik, G., Litany, O., Kasten, Y., & Atzmon, Y. (2026). Joints, camera, action! Co-generating 3D motion and video for arbitrary rigs. arXiv. https://arxiv.org/abs/2610.28632