acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Language-conditioned humanoid motion generation turns a textual request into a sequence of robot poses. A short request leaves timing, support, and posture partly unspecified, yet a useful motion must realize these choices coherently and retain them when the user asks for a local change. We present TRACE, a framework for motion generation and reference-aware editing on the Unitree G1 humanoid robot. Our central idea is to express motion structure and requested revisions through typed kinematic quantities that can be grounded in a motion reference. During training, forward kinematics extracts support, root, and hand events from paired motions to supervise a language model. At inference, the language model predicts duration and event intervals from text; their numerical encodings are aligned to motion patches and combined with the original caption in a diffusion Transformer. A gravity-referenced representation directly predicts height and tilt, while trajectory and contact objectives constrain accumulated motion errors. For revision, a typed request specifies a body–time scope and a goal relative to the current motion. A separate diffusion editor uses source features and revised events to realize that goal, with reference projection preserving protected coordinates throughout sampling. Experiments show strong text–motion alignment and human ratings. Matched ablations support structured event conditioning and improved kinematic consistency, while shared-scope comparisons demonstrate more reliable completion of successive edits than the adapted learned editors.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Humanoid Motion Generation, Conversational Motion Editing, Typed Semantic Planning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_8mvhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

TRACE presents a framework for text2motion generation and editing for the Unitree G1 humanoid. For generation, it extracts confidence-weighted support, root, and hand events from training motions through forward kinematics and supervises a language model to predict corresponding event plans and duration from text. A diffusion Transformer combines numerical, temporally aligned event conditions with the original caption. The system also uses a gravity-referenced motion representation and geometric/contact objectives.

For editing, a separate language adapter and deterministic compiler convert supported requests into body-time scopes and numerical source-relative goals. A separately trained diffusion editor conditions on the source motion, goals, and revised events, while reference projection preserves protected encoded coordinates.

Evaluation includes generation comparisons on a processed BONES-SEED split, matched generation ablations, a human study, qualitative robot examples, and controlled hand-height editing.

Critical strengths
  1. Realization-aligned supervision for intermediate plans is a clean, articulable idea, and the paper tests it honestly. The claim that an intermediate description should be grounded in the specific paired demonstration (because a caption may have many valid realizations) is a real conceptual point that distinguishes this from Motion-R1 / MotionScript / FineMoGen-style caption-derived decomposition. The authors build exactly the matched control that could falsify it and report that it helps on every metric. I value the design of this control more than its magnitude.

  2. Compiling an edit request into a measurable kinematic scope and a metric, source-relative goal, then scoring success by FK rather than by distance to a single "ground-truth" edited motion, is the right evaluation philosophy for motion editing. The multi-turn protocol is more rigorous than what MotionFix-style benchmarks use. If the community adopted "goal-verifiable relative edits" as an evaluation paradigm, that would outlive any particular model, which is why I count it as consequential even though the specific task tested is narrow.

Critical weaknesses

1. The lasting conceptual contribution is insufficiently distinguished from prior work and insufficiently isolated by the generation analysis.

MotionScript already identifies the risk of caption expansion becoming inconsistent with the paired motion and constructs motion-grounded descriptions. CoMo[1], which is not discussed in this submission, already uses motion-derived semantic codes for body-time localization and iterative language-guided editing. TRACE’s numerical goals, continuous diffusion realization, and coordinate projection are meaningful distinctions, but the paper needs to establish what new capability or understanding follows from them. The matched generation ablations support the the proposed conditions, but they do not directly measure the motivating plan–motion conflict, generated-event adherence, or a failure regime that aligned plans resolve. Moreover, Table 5 fixes the same inference planner while changing part of diffusion-training supervision. It therefore does not distinguish the proposed realization-alignment explanation from other changes in training-condition compatibility. My concern is the limited explanatory evidence for the central insight.

2. The controlled editing study does not establish what the proposed learned editor adds beyond existing kinematic tools.

The paper emphasize the ability of whole-body editing, and demonstrates examples like "Jump higher, rest unchanged" (Fig 1). Such tasks requires full-body joints editing or manipulation. However, the actual quantitative benchmark only evaluate a single metric, i.e. single-hand vertical wrist-height offsets on a fixed sine-squared temporal envelope, and explicit time intervals in the request. While other baselines receives different information. Consequently, the success gap does not isolate the contribution of the event representation. The broader whole-body editing example remains qualitative. Together, these limitations prevent the editing results from establishing a substantial advance beyond the restricted task and interface used for training.

References:

[1] Huang, Y., Wan, W., Yang, Y., Callison-Burch, C., Yatskar, M., & Liu, L. (2024, September). Como: Controllable motion generation through language guided pose code editing. In European Conference on Computer Vision (pp. 180-196). Cham: Springer Nature Switzerland.

Questions for authors
  1. Can the authors clarify the precise new insight relative to MotionScript’s motion-grounding argument and CoMo’s semantic-code-based iterative editing? Which reported result most directly establishes the benefit of TRACE’s formulation?
  2. OMG/ECHO/TextOP/TeXedo use official checkpoints trained on different data and are scored by an evaluator trained on your split. Were they retrained on your split, or can you report them on a shared evaluator / shared data? Otherwise the Table 1 rows for these methods mostly measure distribution shift.
  3. How much of the improvement remains when revised source-event conditioning is removed from TRACE? Conversely, how does an adapted learned baseline perform when given the same numerical offset representation? These controls would help identify whether event reuse provides a consequential benefit beyond explicit goal conditioning.
  4. Can the authors quantify the difference between caption-derived plans and paired motions, and show whether generated motions follow their supplied events?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 8mvh

Thank you for asking which experiment shows that the generator executes a plan, and what a learned editor adds once every method receives the same goal.

Do generated motions follow the timing supplied by the plan?

Yes. On segments where foot support changes, supplying event intervals improves overlap with the requested timing by 17.39 percentage points for the left foot and 18.24 for the right, relative to a clip-wide summary. MotionScript grounds descriptions in demonstrations, and CoMo localizes edits with body-part pose codes. TRACE builds on those ideas by predicting when an event should hold and checking the interval against the one extracted from the generated trajectory. A caption can permit several timings. The supplied plan specifies which one to test. Interval IoU measures the overlap between those two intervals.

EventTransition segmentsText-only IoU (%)Clip-wide summary IoU (%)Time-aligned plan IoU (%)Paired gain over summary (pp, 95% CI)
Left-foot support8,87028.4828.4145.81+17.39 [16.95, 17.80]
Right-foot support8,91828.1428.0746.31+18.24 [17.78, 18.69]
Root ascent or descent8,44726.2026.3538.86+12.52 [12.05, 12.97]
Left-hand height3,98229.2830.3936.01+5.62 [5.26, 5.98]
Right-hand height4,14528.8229.5034.26+4.76 [4.44, 5.08]
Root height7537.2736.7041.25+4.56 [2.77, 6.56]

With the inference planner fixed, changing the generator’s training labels improved retrieval and lowered FID.

Inference plannerTraining labelsR@1 (%) ↑R@2 (%) ↑R@3 (%) ↑MM-Dist ↓FID ↓Diversity
SharedCaption-derived73.2685.8990.000.83310.04581.3817
SharedPaired-motion-derived73.8486.2490.310.81650.04341.3766
Trained per label sourceCaption-derived72.9885.8290.010.822400.046101.3757
Trained per label sourcePaired-motion-derived74.0386.5390.620.813780.046381.3732

What does generation add when an edit specifies no joint trajectory?

In the jumping jack, “jump higher, keep the rest unchanged” leaves the trajectory of the pelvis and both legs unspecified. TRACE generated a deeper crouch followed by a higher apex. The user supplied no target height, knee angles or keyframes. At 0.45 s, the pelvis was 9.37 cm lower than in the source and both knees were more flexed. The pelvis then rose 38.29 cm to its apex, compared with 19.35 cm in the source.

Kinematic measureSource referenceTRACE edited referenceChange
Lowest pelvis height before its apex (at 0.45 s in both clips)70.10 cm60.73 cm−9.37 cm
Left / right knee flexion at that instant67.76° / 69.03°98.44° / 98.29°+30.67° / +29.26°
Peak pelvis height (at 0.75 s / 0.80 s, respectively)89.45 cm99.02 cm+9.57 cm
Vertical pelvis travel from the low point to the apex19.35 cm38.29 cm+18.94 cm
Peak clearance of both soles6.24 cm15.56 cm+9.32 cm
Pelvis height at the last frame78.00 cm77.76 cm−0.24 cm

A constant upward shift of 9.57 cm would match the edited apex but would put the pelvis at 79.67 cm at the low point, versus the observed 60.73 cm. In this example, generation provides one coordinated trajectory from the instruction and source motion.

Event features help when the wrist trajectory alone leaves the edit underdetermined. Holding the wrist centre fixed leaves wrist rotation available to move the palm across a requested height boundary.

Palm refinementCategory hits (of 297), seed 1Category hits (of 297), seed 2Height RMSE (cm), seed 1 ↓Height RMSE (cm), seed 2 ↓
Without target-event features2682701.6111.602
With target-event features2872871.6431.645

The whole-body refinement followed instructions on MotionFix. On 1,013 held-out source–target pairs, it came closer to the target on all three measures when it received the paired human-written instruction than when the same instructions were shuffled across sources. The first two columns of scores report gains over copying the source. For paired instructions, the 95% CIs for those gains were 0.00668–0.01806 in target-embedding distance and 0.635–0.853° in mean joint-angle error.

MotionFix conditionReduction in target-embedding distance ↑Reduction in mean angle error over 29 joints (°) ↑Pelvis-relative position error over 16 body sites (cm) ↓
Whole-body refinement, paired instruction0.011480.7328.577
Same refinement, shuffled instruction0.002180.5738.671

Thank you for the additional results. The new experiments showed that generated motions follow the event plans and support the benefit of motion-derived supervision. These provide stronger evidence of my concerns.

I am not an expert in the area but I still find the new insight beyond prior work unclear. The palm experiment does not show how much event conditioning contributes to the original editing results, and the MotionFix comparisons do not establish an advantage over existing editors. I question whether the complex framework could provide substantial impact to the community. I would retain my original score for now.

Reviewer_gFpROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces TRACE, which is a method for text-to-humanoid motion generation and conversational editing. They focus on the Unitree G1 robot. Specifically, the method first uses forward kinematics to extract motion events and then uses them to supervise an LLM for motion planning. Motion generation is then uses these text embeddings and a diffusion Transformer to generate humanoid motions. The authors also propose a gravity-referenced motion representation and additional geometric and contact objectives to improve kinematic consistency. For motion editing, it uses a separate diffusion editor that uses typed instructions, body-time masks, and reference projection to modify selected motion properties. Experiments demonstrate good text-motion alignment and improvements in generation quality compared to existing baselines.

Critical strengths
  1. The proposed motion-grounded event conditioning is very interesting. Rather than relying solely on text features, TRACE extracts structured support, root, and hand events from motion data to provide additional supervision for generation. The ablation studies in Figures 4 and 5 also demonstrates the effectiveness of event conditioning.

  2. The quantitative results are encouraging. In Table 1, TRACE achieves the best R-Precision, MM-Dist, and FID compared to existing methods, including MotionGPT3 and MDM.

  3. Support for conversational motion editing is also a good extension.

Critical weaknesses
  1. Although conversational motion editing is listed as one of the central contributions, the quantitative editing experiments are mainly limited to hand-height adjustments of 5 cm. It is currently missing quantitative evaluation of many complex editing tasks, such as changing whole-body movements, foot contacts, or motion trajectories (only appears in qualitative results). This makes it difficult to assess whether the proposed claims are well supported.

  2. The paper introduces multiple components for motion editing, including typed instructions, revised event conditions, a source-conditioned diffusion editor, and reference projection. It remains unclear how much each proposed component contributes to the improvement.

  3. The paper mainly compares against relatively established diffusion methods such as MDM, while overlooking many recent advances, including MotionDiffuse, MLD, StableMoFusion, MotionLCM, MARDM, SALAD, and MotionHiFlow, etc. Some of these methods also address motion consistency, fine-grained control, and editing, which are closely related to the proposed framework. I would encourage the authors to provide a more comprehensive discussion of these studies and, if possible, include stronger recent diffusion-based baselines rather than relying primarily on MDM.

  4. From a reviewer's perspective, TRACE mainly extends existing text-to-motion generation techniques to the Unitree G1 humanoid setting. The framework combines LLM-based planning, diffusion generation, kinematic losses, and reference-preserving editing, many of which have been explored individually in previous works. In particular, MotionLCM V2, MaskControl, OmniControl, ACMDM, and ProjFlow have already investigated fine-grained motion control, editing, or constraint-based generation. However, these relevant approaches are not considered in the experimental comparisons.

  5. TRACE introduces a gravity-referenced representation and several kinematic objectives to improve motion consistency. Although Figure 5 provides ablations showing improvements in selected kinematic metrics, it remains unclear whether these relatively complex designs are necessary compared with simpler alternative representations and loss formulations.

Questions for authors

Please see the weakness section and address accordingly.

3 comments
AuthorsOpenReview ↗

Response to Reviewer gFpR

Thank you for asking us to compare recent generators on BONES-SEED and test edits beyond the wrist.

How does TRACE compare with recent BONES-SEED generators?

TRACE ranked first on all five generation metrics against the eight added BONES-SEED configurations. TRACE exceeded HiFlow, the strongest R@1 baseline, by 2.96 percentage points (73.84% versus 70.88%). All methods received the same predicted lengths and used the same frozen evaluator.

GeneratorR@1 (%) ↑R@2 (%) ↑R@3 (%) ↑MM-Dist ↓FID ↓
TRACE73.8486.2490.310.81650.0434
HiFlow70.8881.3884.900.83690.0519
SALAD70.8682.7986.740.84520.0495
StableMoFusion69.6781.1684.960.88510.0931
MARDM65.2877.6482.850.88060.0604
MotionLCM teacher64.7478.1383.580.87870.0842
MotionLCM student63.6077.4282.830.89670.0908
MotionDiffuse56.0870.0476.180.96350.0635
MLD54.3068.0273.960.97670.0684
The eight configurations cover seven recent methods. Baseline generators generally received 50,000 training updates, latent encoders or autoencoders received 40,000 where applicable, and the MotionLCM student received 20,000 additional distillation updates. Validation data governed model selection.

How well do whole-body edits preserve the source motion?

The paired instruction improved target matching on all three MotionFix measures. We evaluated a separate source-conditioned refinement on 1,013 held-out MotionFix pairs. Compared with copying the source, refinement reduced mean joint-angle error by 0.732° (95% CI 0.635–0.853°), reduced target-embedding distance by 0.01148 (95% CI 0.00668–0.01806), and lowered pelvis-relative position error from 8.760 to 8.577 cm. Shuffling the instructions weakened all three gains.

MotionFix conditionReduction in target-embedding distance ↑Reduction in mean angle error over 29 joints (°) ↑Pelvis-relative position error over 16 body sites (cm) ↓
Copy source motion008.760
Whole-body refinement, paired instruction0.011480.7328.577
Same refinement, shuffled instruction0.002180.5738.671
All three methods preserved the protected waist and arm angles, and TRACE gained the most clearance on this reference motion. The three methods edited one archived jumping jack with a common mask. We measured peak clearance of both soles and the maximum change in 17 protected joint angles. The single TRACE edit gained 9.32 cm, versus three-seed means of +3.15 cm for TMED-G1 and −0.37 cm for SimMotionEdit-G1.
MethodChange in peak clearance of both solesMaximum change across 17 protected joint angles
TRACE+9.32 cm< 0.00001°
TMED-G1+3.15 cm< 0.00001°
SimMotionEdit-G1−0.37 cm< 0.00001°

How does TRACE compare with existing spatial control methods?

We compared ProjFlow, OmniControl, ACMDM, MotionLCM V2 and MaskControl on the dense trajectory constraints each supports. ACMDM’s single-joint interface covers the root task. These methods already demonstrate fine-grained control. The specific editing question for TRACE is whether it can apply a change to a supplied motion while preserving the user’s protected joint angles.

Which components improve editing success?

Source-motion features improved four-turn editing success by 20.63 percentage points, and parser fine-tuning improved end-to-end success by 4.52 points.

Controlled changeMeasured outcomeGain (percentage points)95% CI for gain
Enable explicit source-motion featuresFour-turn success, 6.31% → 26.94%+20.6316.69 to 24.62
Fine-tune the typed parserEnd-to-end success+4.522.01 to 7.54

Which representation and loss choices improve foot contact?

The contact objective produced the largest reduction in foot speed in this sequence of ablations. We inferred foot contacts from each generated motion. Gravity-based coordinates alone reduced sliding speed from 0.4146 to 0.3941 m/s. Adding the geometry objective increased it to 0.4201 m/s. Adding the contact objective then brought it down to 0.3587 m/s and yielded the lowest mean penetration (0.1583 cm).

ConfigurationSliding speed during support (m/s) ↓Mean floor penetration (cm) ↓Contact-point coverage (%)Clips without valid contact (%)
Heading-local coordinates0.41460.262868.260.023
Gravity-based representation0.39410.174474.660.098
Add geometry objective0.42010.210274.930.165
Add contact objective (full model)0.35870.158375.440.203

Dear reviewer, the authors have responded to your comment. Could you please take a look? Let us know if you have further concerns.

Dear Reviewer,

As we approach the end of the discussion period, we hope you have had time to review our rebuttal and revised manuscript. If our responses have adequately addressed your concerns, we would sincerely appreciate it if you could consider updating your rating. Otherwise, we would be grateful for any further feedback on the remaining issues.

Thank you very much for your time and thoughtful feedback.

Reviewer_DhMXOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses text-to-motion generation and reference-aware editing for the Unitree G1 humanoid. A frozen language model predicts a compact, typed plan (support, root, and hand events) grounded in motion-derived quantities rather than emitting free-form edits, and a diffusion Transformer generates or revises the motion with per-step reference projection to preserve protected coordinates. Generation is evaluated on the BONES-SEED split against eight baselines using automatic metrics and a user study, while editing is measured on a controlled hand-height protocol. The paper also reports real-robot deployment and ablation studies.

Critical strengths
  1. Tying the edit vocabulary to quantities that can be computed from the trajectory (support, root height, palm height) and then enforcing preservation through per-step reference projection is the right way to keep an LLM from inventing unsupported motion. The shared-mask construction gives a preservation guarantee that can be checked directly — the method produces zero protected-coordinate errors by construction — and the scope diagnostics (Fig. 6) show that body selection, temporal localization, and kinematic-chain expansion are each doing real work.
  2. TRACE improves R@1 from 64.46 to 73.84 over the strongest baseline (MotionGPT3), with matching gains in MM-Dist (0.8886→0.8165), FID (0.0628→0.0434), and a large jump in the user study (Overall 1.419→2.391 on a 3-point scale). Automatic and human metrics agreeing this closely is rare and is the main reason I lean positive. The matched ablations (Tables 4–5, Fig. 4), which hold the entire system fixed and vary only the plan, are exactly the control one wants.
Critical weaknesses
  1. All quantitative editing numbers come from a single controlled task: wrist-height edits with a fixed 5 cm signed offset following a fixed sine-squared profile over an explicitly specified interval, using scripted English requests over 199 sources. The whole-body, free-form edit that motivates the paper (Fig. 1, "jump higher, legs unchanged") is explicitly excluded from the quantitative cohort. On the controlled protocol, all-turn success at depth 4 is only 30.82% (Fig. 7) — fewer than one in three four-turn chains fully meets the change-and-preservation criteria. The roughly 5–11× margin over the two adapted baselines is real, but it does not establish competence on the unscripted, multi-intent, whole-body conversational edits named in the title and abstract.
  2. Every number is on BONES-SEED, which the authors curate and process with an extensive, partly heuristic pipeline (including Qwen-Max semantic classification and a long sequence of filters and retargeting steps), rather than on a standard benchmark such as HumanML3D/AMASS. Most comparators are 2026 arXiv/ICLR papers from the same ecosystem. The results therefore cannot be placed against the broader text-to-motion literature, and they inherit the authors' retargeting and tuning choices. The user study, though strongly favorable, uses 100 prompts, reports large SDs (±0.40 for TRACE), and gives no significance test.
  3. In the matched control the aligned plan adds only +0.74 R@1 (absolute) over text-only (73.10→73.84, Table 4); the large gap to MotionGPT3 is mostly carried by the bundled gravity/contact representation and the data pipeline. The gravity-gauged state itself is height plus two rotation-matrix columns, and its kinematic-error reduction is modest (Fig. 5). The components are useful, but the specifically novel element is the smallest measured contributor.
Questions for authors

1.Can you report editing success on genuinely free-form, unscripted requests — including paraphrases, omitted intervals, and whole-body edits beyond wrist height? Of the failures behind the 30.82% all-turn number, how many are goal-achievement failures versus preservation failures? 2.What is the parse/compile failure rate of the frozen planner on held-out user language (ambiguous, negated, or out-of-range requests)? The 199/200 naturalness figure describes the frozen candidate set, not incoming user language. 3.Are the user-study differences vs. MotionGPT3 statistically significant? Please report per-rater paired comparisons or bootstrap CIs rather than only means and SDs across 16 raters. 4.Can you provide results on a standard benchmark (HumanML3D), or run the canonical baselines from their published checkpoints, so the generation gains can be assessed outside BONES-SEED? Please also clarify the provenance relationship among BONES-SEED, MotionGPT3, and the authors, and who tuned each baseline.

3 comments
AuthorsOpenReview ↗

Response to Reviewer DhMX

Thank you for suggesting the HumanML3D comparison.

How well do scripted and free-form edits work?

The human-written instruction helped on single-turn whole-body edits. We evaluated a separate whole-body refinement on 1,013 held-out MotionFix pairs. Compared with copying the source, the paired instruction reduced target-embedding distance by 0.01148 and mean joint-angle error by 0.732°. Shuffling the instructions reduced those gains to 0.00218 and 0.573°.

MotionFix instructionReduction in target-embedding distance ↑Reduction in mean angle error over 29 joints (°) ↑
Paired instruction0.011480.732
Shuffled instruction0.002180.573
In the single “jump higher” case, TRACE increased peak clearance of both soles by 9.32 cm while preserving 17 waist and arm joint angles.

Do the generation results transfer to HumanML3D?

TRACE ranked first on all five metrics among the four main methods retrained on HumanML3D. We evaluated all four on the same 4,496 requests with common lengths, motion representation, evaluator and retrieval groups. TRACE’s R@1 exceeded MotionGPT3’s by 3.97 percentage points (54.80% versus 50.83%), and its FID was 0.1001, compared with 0.2286 for MoMask, the closest baseline on that metric.

MethodR@1 (%) ↑R@2 (%) ↑R@3 (%) ↑MM-Dist ↓FID ↓
TRACE54.8074.6383.762.78290.1001
MotionGPT350.8370.8080.973.03500.9710
MoMask48.3267.5277.413.21370.2286
MDM33.7451.9363.564.00901.0636

The authors have no affiliation with BONES-SEED or MotionGPT3.

Which generator did the raters prefer, and how certain is the estimate?

Raters preferred the earlier TRACE generator to MotionGPT3. Sixteen raters scored the motions on a three-point scale. The mean advantage was 0.97 points (paired 95% CI 0.70–1.22, resampling raters and prompts).

How much of the gain comes from the time-aligned plan?

Time alignment reduced MM-Dist by 0.0145 and FID by 0.0014 relative to a clip-wide event summary. On the same 13,302 BONES-SEED requests, text-only conditioning yielded a mean text–motion embedding distance (MM-Dist) of 0.8341 and an FID of 0.0464. The clip-wide summary lowered these to 0.8310 and 0.0448, and the time-aligned plan lowered them further to 0.8165 and 0.0434. Model size, predicted duration and training budget were held fixed.

Dear reviewer, the authors have responded to your comment. Could you please take a look? Let us know if you have further concerns.

Dear Reviewer,

As we approach the end of the discussion period, we hope you have had time to review our rebuttal and revised manuscript. If our responses have adequately addressed your concerns, we would sincerely appreciate it if you could consider updating your rating. Otherwise, we would be grateful for any further feedback on the remaining issues.

Thank you very much for your time and thoughtful feedback.

Reviewer_XpADOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Typed Reference-Aware Conversational Editing (TRACE), a framework for generating and editing motion for the Unitree G1 humanoid from language instructions. Specifically, the authors introduce a motion-grounded planner that predicts duration and typed support, root, and hand events to guide a diffusion Transformer alongside the original text. A gravity-referenced representation and geometric and contact objectives further encourage kinematic consistency. For editing, follow-up requests are compiled into a body-time scope and a goal relative to the current motion. A separately trained diffusion editor uses these conditions and revised source events to make the requested change, while reference projection preserves protected coordinates. Experiments on BONES-SEED show improvements in text-motion alignment, perceived quality, and edit completion, with additional qualitative demonstrations on a real robot.

Critical strengths
  1. Grounding the intermediate plan in paired motion gives its support and height events a clear kinematic interpretation. The matched controls in Tables 4 and 5 show consistent, though modest, gains from time-aligned events and motion-derived supervision.

  2. The complete generator performs well in both automatic and human evaluations. In Table 1, it improves R@1 from 68.46 to 73.84 over MotionGPT3 and receives higher ratings for motion quality and agreement with the text.

Critical weaknesses
  1. The plan is intended to provide explicit temporal guidance, but its evaluation mainly measures overall generation quality. Without showing whether generated motions follow the predicted support and height intervals, it is difficult to determine how the proposed structure is being used.

  2. The editing comparison combines revised event conditioning with direct numerical wrist targets, while the adapted baselines retain their text-and-source interfaces. The higher success rate therefore supports the complete editor but does not establish what the event-based design adds beyond explicit goal conditioning. Besides, sliding, drift, and contact-height errors are measured over the reference motion’s support intervals, which can penalize a valid motion with different timing.

Questions for authors
  1. The proposed event Plan is intended to capture temporal structure that is missing from natural-language instructions. However, the current ablations mainly evaluate overall generation quality. How accurately do the generated motions follow the predicted support, root-height, and hand-height events, particularly when a caption admits multiple valid motion sequences?

  2. The gravity-referenced representation and contact objectives are designed to improve kinematic consistency, but several physical metrics rely on contact intervals from the reference motion. How do the authors distinguish genuine improvements in physical consistency from closer agreement with the reference contact timing? Do the improvements remain when contacts are determined from the generated motion itself?

  3. The editing framework combines explicit relative goals, revised motion events, and reference-conditioned diffusion. However, the reported comparisons do not establish the individual contribution of these design choices. How does the event-aware editor compare with its relative-goal-only counterpart and the bounded inverse kinematics procedure used to construct its training targets?

  4. The paper motivates conversational editing through requests such as "Jump higher, rest unchanged," but the quantitative evaluation is restricted to local hand-height adjustments. How does the proposed editing framework handle more complex requests involving coordinated changes across multiple body parts, particularly when root motion affects the world-space positions of otherwise protected regions?

2 comments
AuthorsOpenReview ↗

Response to Reviewer XpAD

Thank you for asking how we verify that a motion follows the requested timing and what stays fixed when a user asks for a higher jump.

How closely does the motion follow a supplied timing plan?

The time-aligned plan improved overlap with the requested intervals for all six event types. We scored each generated motion against its supplied plan, since a caption may admit several valid timings. Interval IoU measures the intersection over union of the supplied and generated intervals. Compared with a clip-wide summary, the largest gains were for left- and right-foot support (+17.39 and +18.24 percentage points). The root-height gain (+4.56 points).

EventTransition segmentsText-only IoU (%)Clip-wide summary IoU (%)Time-aligned plan IoU (%)Paired gain over summary (pp, 95% CI)
Left-foot support8,87028.4828.4145.81+17.39 [16.95, 17.80]
Right-foot support8,91828.1428.0746.31+18.24 [17.78, 18.69]
Root ascent or descent8,44726.2026.3538.86+12.52 [12.05, 12.97]
Left-hand height3,98229.2830.3936.01+5.62 [5.26, 5.98]
Right-hand height4,14528.8229.5034.26+4.76 [4.44, 5.08]
Root height7537.2736.7041.25+4.56 [2.77, 6.56]

Do the contact gains survive changes in footfall timing?

The full model reduced sliding and penetration even when each output determined its own contact timing. Across 13,302 generated motions, we used each motion’s foot height to identify its support intervals. Relative to heading-local coordinates, sliding speed fell from 0.4146 to 0.3587 m/s (paired 95% CI for the reduction, 0.0520–0.0604 m/s), and mean penetration fell from 0.2628 to 0.1583 cm.

ConfigurationSliding speed during support (m/s) ↓Mean floor penetration (cm) ↓Contact-point coverage (%)Clips without valid contact (%)
Heading-local coordinates0.41460.262868.260.023
Gravity-based representation0.39410.174474.660.098
Add geometry objective0.42010.210274.930.165
Add contact objective (full model)0.35870.158375.440.203

What do the numerical goal, source motion, and event features each contribute to editing?

A numerical target helped both learned editors For TRACE this was a gain of 15.10 percentage points over its 11.84% text-and-source result (paired 95% CI 11.58–18.73). Event features helped the palm cross the requested height boundary. With the wrist trajectory fixed, the network used wrist rotation to change palm height. Both training seeds reached the target category in 287 of 297 cases with event features, up from 268 and 270 without them.

Can a coordinated whole-body edit preserve the protected joints?

All three methods preserved the 17 protected joint angles to within 0.00001°. TRACE produced the largest clearance gain on this example. All methods edited the same jumping jack with the same mask. TRACE increased peak clearance of both soles from 6.24 to 15.56 cm (+9.32 cm). The three-seed mean changes were +3.15 cm for TMED-G1 and −0.37 cm for SimMotionEdit-G1.

MethodChange in peak clearance of both solesMaximum change across 17 protected joint angles
TRACE+9.32 cm< 0.00001°
TMED-G1+3.15 cm< 0.00001°
SimMotionEdit-G1−0.37 cm< 0.00001°

Using the paired instruction brought whole-body edits closer to the target than either copying the source or shuffling the instructions. On 1,013 held-out MotionFix pairs with human-written instructions, the separate whole-body refinement reduced mean error across 29 joint angles by 0.732° (95% CI 0.635–0.853°) and target-embedding distance by 0.01148 (95% CI 0.00668–0.01806) relative to copying. Shuffling the instructions reduced the embedding gain to 0.00218.

MotionFix conditionReduction in target-embedding distance ↑Reduction in mean angle error over 29 joints (°) ↑Pelvis-relative position error over 16 body sites (cm) ↓
Copy source motion008.760
Whole-body refinement, paired instruction0.011480.7328.577
Same refinement, shuffled instruction0.002180.5738.671
Reviewer_XpADOpenReview ↗

Official Comment by Reviewer XpAD

Thank you for the additional results. The new experiments provide stronger evidence for the effectiveness of event-based planning and the improvements in motion consistency. However, I still find it unclear what the event representation adds to conversational editing beyond explicit numerical goals and reference projection. The additional results also do not sufficiently establish the broader whole-body editing capability claimed by the paper. I therefore remain unconvinced that the paper meets the bar for acceptance, but I will raise my score from 1 to 2 while maintaining my original low confidence level.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.