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

The gist The central challenge is to connect an underspecified request to concrete motion structure that can subsequently be revised relative to an accepted reference.

How it works

TRACE is a generation and editing framework for the Unitree G1 humanoid robot built around typed, measurable kinematic quantities The core insight is that a language model learns to predict support, root, and hand events from motion-derived supervision; their time-aligned encodings complement the original text in a diffusion generator Gravity-referenced states and trajectory–contact objectives improve kinematic consistency For editing, typed requests define body–time scopes and source-relative goals. A separate reference-conditioned diffusion editor uses these goals and revised events to realize changes, while sampling-time projection preserves protected state coordinates.

Motion-Grounded Planning and Conditioning

The system constructs a compact event vocabulary from measurable properties of the paired training motion. This motion-derived supervision provides realization-aligned supervision for the language-to-plan predictor. Specifically, Each four-frame patch carries two independent support probabilities and four categorical distributions. The plan prediction is conditioned using a mixture of motion targets and predicted events, formulated as C mix = W ⊙ Y + (1 − W) ⊙ h(u, N), with probability 1/2, h(u, N), otherwise.

Gravity-Gauged and Contact-Consistent Diffusion

To ensure kinematic consistency, the system represents height and tilt relative to gravity and supervises integrated geometry and foot contact explicitly. The state representation is a 40D motion representation that includes absolute height and gravity-relative tilt. The generator's objective is organized into three weighted groups: Lgen = Lmotion + Lgeometry + Lcontact. Contact losses include specific terms such as Lslip = D ρβv squared / 2(p f,xy t+1,k - p f,xy t,k) and Lheight = D ρβh(z f t − z¯gt) E c f t.

Automatic Masking and Reference-Aware Motion Editing

Editing involves three key components: automatic body–time masking, source-relative goals, and reference projection. The Automatic scope localization uses a supervised LLM to predict an attribute, effector, signed change, and interval to construct the body–time mask St(f, j) = 1[j ∈ K(et)] 1[at ≤ f < bt]. The Reference-preserving transitions use a stochastic state transition Mt+1 ∼ Tη(· Xt, At) = pη(· f(ut), Mt, St, Dt) where the binary mask S preserves protected normalized coordinates by reference projection.

Evaluation and Results

TRACE leads compared methods in retrieval, MM-Dist, FID, and human ratings. The experiment on editing shows that TRACE achieves higher all-turn success than the adapted learned editors under shared scopes and reference projection. In the current-cohort editing comparison, TRACE achieved a success rate of 30.82% for four turns, compared to 6.48% for TMED-G1.

The paper concludes that Typed motion quantities thus connect language-conditioned generation with measurable revisions of an evolving reference. The system addresses text instructions and flat-ground kinematics without scene perception. All four errors are converted to centimeters. The final step evaluates Contact as a bundle.

--- Page 9

The gist The central challenge is to connect an underspecified request to concrete motion structure that can subsequently be revised relative to an accepted reference.

How it works

TRACE is a generation and editing framework for the Unitree G1 humanoid robot built around typed, measurable kinematic quantities The core insight is that a language model learns to predict support, root, and hand events from motion-derived supervision; their time-aligned encodings complement the original text in a diffusion generator. Gravity-referenced states and trajectory–contact objectives improve kinematic consistency For editing, typed requests define body–time scopes and source-relative goals. A separate reference-conditioned diffusion editor uses these goals and revised events to realize changes, while sampling-time projection preserves protected state coordinates.

Motion-Grounded Planning and Conditioning

The system constructs a compact event vocabulary from measurable properties of the paired training motion. This motion-derived supervision provides realization-aligned supervision for the language-to-plan predictor. Specifically, Each four-frame patch carries two independent support probabilities and four categorical distributions. The plan prediction is conditioned using a mixture of motion targets and predicted events, formulated as C mix = W ⊙ Y + (1 − W) ⊙ h(u, N), with probability 1/2, h(u, N).

Gravity-Gauged and Contact-Consistent Diffusion

To ensure kinematic consistency, the system represents height and tilt relative to gravity and supervises integrated geometry and foot contact explicitly. The state representation is a 40D motion representation that includes absolute height and gravity-relative tilt The generator's objective is organized into three weighted groups: Lgen = Lmotion + Lgeometry + Lcontact. Contact losses include specific terms such as Lslip = D ρβv squared / 2(p f,xy t+1,k - p f,xy t,k) and Lheight = D ρβh(z f t − z¯gt) E c f t.

Evaluation and Results

TRACE leads compared methods in retrieval, MM-Dist, FID, and human ratings The experiment on editing shows that TRACE achieves higher all-turn success than the adapted learned editors under shared scopes and reference projection. In the current-cohort editing comparison, TRACE achieved a success rate of 30.82% for four turns, compared to 6.48% for TMED-G1.

1. Improved motion generation via grounded conditioning: The system can now ground generation conditions in paired training motions, allowing a language model to predict event targets (support, root, hand events) from motion-derived supervision to create a motion-grounded sparse-event plan.

2. Enhanced kinematic consistency through gravity referencing: The diffusion model incorporates gravity-referenced states and trajectory–contact objectives by directly representing absolute height and gravity-relative tilt, which helps reduce accumulated trajectory errors, as the objective groups are defined as Lgen = Lmotion + Lgeometry + Lcontact.

3. Reliable interactive refinement via reference awareness: The system can handle successive edits where a request is interpreted against the current motion, using reference projection preserves protected coordinates during editing, ensuring that changes are realized while retaining the rest of the encoded motion.

4. Superior multi-turn editing performance: By compiling requests into a body–time scope and a source-relative goal, the system can execute successive relative requests under shared preservation constraints, achieving higher all-turn success compared to adapted learned editors.

5. Enhanced semantic alignment: The combination of motion-grounded conditioning and kinematic consistency leads to improved semantic alignment and motion quality over the compared methods in generation, as demonstrated by superior retrieval, MM-Dist, and FID scores in experiments.

Language-conditioned humanoid motion generation enables users to create and refine robot movements through natural language. The central challenge is to connect an underspecified request to concrete motion structure that can subsequently be revised relative to an accepted reference. We present TRACE, a generation and editing framework for the Unitree G1 humanoid robot built around typed, measurable kinematic quantities. Our insight is to ground generation conditions in paired training motions and editing goals in the current motion. A language model learns to predict support, root, and hand events from motion-derived supervision; their time-aligned encodings complement the original text in a diffusion generator. Gravity-referenced states and trajectory--contact objectives improve kinematic consistency. For editing, typed requests define body--time scopes and source-relative goals. A separate reference-conditioned diffusion editor uses these goals and revised events to realize changes, while sampling-time projection preserves protected state coordinates. Quantitative evaluation and human assessment show improved semantic alignment and motion quality over the compared methods. Controlled ablations support the value of motion-grounded conditioning and kinematic consistency, while multi-turn editing evaluations demonstrate more reliable execution of successive relative requests under shared preservation constraints. Together, these results support typed kinematic structure as a common basis for motion generation and reference-aware revision.