TRACE: Typed Reference-Aware Conversational Editing for Humanoid Motion Diffusion

Can a humanoid motion generator be made both more faithful to text and editable in conversation, by predicting measurable events such as foot support, root height and hand height and then using them to scope later changes?

A caption rarely says when a foot supports the body or how high a hand rises, so text-to-motion models fill in structure that may not match the training motion. Editing is harder still: a relative request such as 'raise the hand' has to be interpreted against the current motion while everything else stays as it was, and repeated edits tend to drift.

Text-to-motion models (MDM, MoMask, MotionGPT3) map captions to trajectories or motion tokens; humanoid-specific generators (TextOp, TeXedo, ECHO, OMG) target the G1 directly. Language-model planners break instructions into detailed descriptions, and source-conditioned editors (TMED, SimMotionEdit) learn changes from text plus an existing sequence, without a typed description of what should change.

A language model learns to predict a sparse event plan (left and right foot support, root vertical speed and height, left and right hand height in bins) from labels computed on the paired training motion, not from the caption. Time-aligned encodings of that plan join the text in a diffusion generator that uses gravity-referenced states plus trajectory and foot-contact objectives. For editing, a typed request defines which body part and time window may change and a goal relative to the source; a separate reference-conditioned diffusion editor makes the change, and projection during sampling keeps protected coordinates fixed.

On the BONES-SEED test set retargeted to the G1, TRACE reached R@1 of 73.84%, R@2 86.24% and R@3 90.31%, against 68.46%, 81.29% and 85.81% for MotionGPT3; FID fell to 0.0434 from 0.0628 and MM-Dist to 0.8165 from 0.8886. In a user study on 100 prompts with 16 participants, TRACE scored 2.338 for quality, 2.444 for text match and 2.391 overall on a 1 to 3 scale; the best rival overall was ECHO at 1.469. Editing was tested on 199 held-out G1 motions with 5 cm hand-height requests chained into 597 four-turn sequences, where the authors report more reliable all-turn success than adapted TMED and SimMotionEdit (results shown in figures).

This is a preprint. The system handles text and flat-ground kinematics with no scene perception, and the authors note that contact objectives reduce but do not eliminate foot slip. Metrics assess generated motion, not whether each predicted event is obeyed exactly. Real-robot evidence is limited to qualitative demonstrations in an appendix, with no success rate on hardware, and editing was tested only on hand-height changes.