{"id":"b41f1216-04d8-4a91-ad54-0370a8b37c03","arxiv_id":"2501.16557","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CARING-AI combines ChatGPT text generation, environment scanning, and smoothed text-to-motion diffusion to let authors create spatially grounded AR avatar instructions without coding or motion capture.","lead":"CARING-AI is an augmented reality authoring system that generates context-aware humanoid avatar instructions from spoken text and a walk-through scan of the environment. The paper shows that authors can create AR instructions without coding or motion capture, with user studies reporting lower workload and faster authoring than a physical demonstration baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Context-awareness claim rests on unmeasured perception accuracy; hand-object interaction is acknowledged subpar, so spatial grounding is unverified.","rationale":"The paper is an HCI systems paper; its central contribution is the workflow for authoring context-aware AR instructions. The strongest discriminator from plain text-to-motion is the claim that the generated avatar instructions are spatially grounded in the physical environment. That grounding is implemented via a perception pipeline (YOLO + MegaPose) whose output is used both to overlay virtual objects and to seed the hand-object motion generation. The paper provides no accuracy evaluation for this pipeline in the deployment environments, and the authors' own §5.3 and User Study observations indicate hand-object interactions are a weak point. I found this more load-bearing than the temporal-smoothing quantitative issue: even if the 0.03 m transition distance were an artifact of circular metric design, the authoring workflow could still produce acceptable instructions; but if spatial grounding fails, the 'context-aware' claim itself collapses. The proposed test would settle the concern by quantifying perception accuracy and the degree to which manual corrections were needed. The reader's verdict of CONDITIONAL remains appropriate: the paper is promising but the context-awareness claim needs this evidence.","tokens_in":31166,"tokens_out":7962,"duration_ms":76594,"concrete_test":"Re-run the perception stack on the recorded HoloLens screenshots from User Studies 1 and 2 scenes (kitchen, living room) and compare against ground-truth object poses (e.g., AprilTag markers or manual annotation). Report YOLO detection recall, MegaPose ADD error, and the count/frequency of manual 6DoF adjustments per participant. Then measure the avatar hand-to-object distance at interaction frames in the final generated animations. If detection recall is high (>90%), ADD error is below 2 cm, manual corrections are rare (e.g., <20% of objects), and hand-to-object distance is below a plausible threshold, the context-awareness claim holds; otherwise it requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that CARING-AI 'blends in the context spatially' for hand-object instructions depends on the local-spatial grounding pipeline (§4.4.2): fine-tuned YOLO detection (§4.6) and MegaPose 6DoF from HoloLens screenshots, which feed object poses used to overlay virtual objects and guide hand-object motion generation. No detection recall, pose error, or manual-correction frequency is reported in the kitchen/living-room study environments. The authors themselves state in §5.3 that 'the quality of hand motion generated by both GMD and our method is subpar... the pickup motion does not adequately display grabbing gestures,' and User Study 1 records P12's confusion about the hand 'automatically sticks to the hand.' Because the system includes an explicit user step to adjust 6DoF (§6.1), a missed object or wrong pose does not categorically break authoring, but it converts the 'natural contextual information' claim into an unmeasured manual-alignment burden. Without accuracy numbers, we cannot tell whether the avatar-object alignment observed in the user study came from perception or from manual fixes; the context-awareness claim is therefore not yet supported quantitatively.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"CARING-AI is an AR authoring system that combines an LLM (ChatGPT) for step-by-step text instructions, a HoloLens 2 scan mode that records user trajectories and runs YOLO-based object detection plus MegaPose 6DoF pose estimation, and a modified Motion Diffusion Model (MDM/GMD) that generates humanoid avatar motion conditioned on text, trajectories, and object locations. A temporal smoothing stage blends transition segments to produce longer seamless animations. The paper reports a design space (spatial/temporal context by local/global content), three application scenarios (asynchronous, remote, ad hoc), a quantitative comparison against GMD on 10 tasks, and two user studies (N=12 each): one on usability of the system and one within-subject comparison with a video-based Programming-by-Demonstration baseline. The central claim is that CARING-AI enables code-less, mocap-free authoring of context-aware AR avatar instructions that are spatially grounded and temporally continuous.","tokens_in":31406,"tokens_out":6559,"duration_ms":64879,"significance":"The contribution is timely and addresses a real gap: state-of-the-art text-to-motion models produce animations that are not situated in a physical workspace, and prior AR authoring tools require programming or motion capture. The design-space decomposition and the asynchronous/remote/ad hoc scenario demonstrations provide a useful framework for AIGC-in-AR research. The authors evaluate against an external baseline (GMD) and external data (HumanML3D), borrow the 0.1 m plausibility threshold from GMD rather than fitting it, and use standard instruments (SUS, NASA-TLX) in the user studies. They also report limitations candidly, including subpar hand motion, rigid-object-only interactions, and HoloLens FOV constraints. If the central claims are substantiated, the system would be a practical step toward democratizing AR instruction authoring. However, the quantitative support is thin in several load-bearing places, particularly the statistical reporting of the smoothing result, the unmeasured perception accuracy behind spatial grounding, and the multiple-comparison issues in User Study 2.","major_comments":[{"comment":"The central quantitative claim that temporal smoothing reduces transition distance to 0.03 m (p<0.05) is reported as two aggregate means across 10 tasks, with no variance, no per-scenario breakdown, no number of generated sequences or transitions, and no value for the smoothing length L in Eq. (1). The choice of 90 frames per instruction and the training/masking details of the modified diffusion model are also not specified. As written, the p-value cannot be checked and the magnitude of the effect relative to variance is unknown; the claim that discontinuities are \"eliminated\" is stronger than the metric supports. Please provide per-scenario results, sample sizes, effect sizes or confidence intervals, and the exact hyperparameters (L, frame counts, seeds) needed for reproduction.","section":"§5.2–5.3, Table 2"},{"comment":"The spatial context-awareness claim rests on the local-spatial grounding pipeline: fine-tuned YOLO detection, MegaPose 6DoF estimation on HoloLens RGB frames, and overlay of virtual objects. The paper reports no detection recall, pose error, or frequency of manual 6DoF correction in the kitchen/living-room environments, even though §6.1 includes an explicit user step to adjust misaligned virtual objects. Since the authors themselves state in §5.3 that hand-object motion quality is subpar and User Study 1 records P12's confusion about the hand \"automatically sticking\" to the object, the quantitative support for \"blending in the context spatially\" is incomplete. In addition, Eq. (4) measures the root-joint distance to a keypoint, not hand-object alignment, so it does not substantiate local spatial grounding. Please report perception accuracy on the study scenes or an ablation separating automatic grounding from manual alignment, and report a hand-object alignment metric.","section":"§4.3–4.4.2, §4.6, §5.3"},{"comment":"The temporal smoothing function is a per-joint convex combination (sigmoid) followed by linear-interpolation resampling. Minimizing the joint-position distance between the last frame of one action and the first frame of the next does not guarantee physical plausibility of the interpolated frames: blended poses can exhibit foot sliding, global drift, or violation of joint-angle limits, all of which are unmeasured. The manuscript's claim that the result is \"fluid,\" \"connected,\" and free of discontinuity requires either a velocity/acceleration-continuity metric, a kinematic-plausibility check, or a perceptual evaluation; a single scalar transition distance is not sufficient.","section":"§4.4.3, Eq. (1)–(2)"},{"comment":"User Study 2 performs multiple paired comparisons without correction for multiple testing: the NASA-TLX subscales, error rate, time, and five Likert items are each tested at α=0.05, and several p-values are near the threshold (0.025–0.046). With N=12, the family-wise error rate is material. The PbD baseline is a home-built video-to-3D pipeline requiring manual segmentation and camera calibration, and no comparison with mature MoCap or video-based authoring systems is provided. The comparative conclusion should therefore be framed as \"vs. this particular PbD setup\" and supported with effect sizes and corrected p-values (or a pre-registered primary outcome).","section":"§7.2"}],"minor_comments":[{"comment":"The update reads \\tilde{K}_t = \\alpha_t K_t + (1-\\alpha_t) K_t, which is algebraically K_t; the two operands should be K^1_t and K^2_t as in Eq. (1). Please fix.","section":"Algorithm 1, line 2"},{"comment":"The sigmoid expression is typeset incorrectly (missing the closing parenthesis in the exponent, '−( 𝑡−( 𝐿/2)'); as written it is not a well-formed function.","section":"Eq. (1)"},{"comment":"There are several typographical errors: \"compared wt\" should be \"compared with,\" \"disconuity\" should be \"discontinuity,\" and \"A vatar\" should be \"Avatar.\" The title's capitalized \"INstruction\" also appears nonstandard.","section":"§7, §5.3, §6.2.1"},{"comment":"The task list contains duplicated and inconsistent entries: \"Closing a Window\" appears twice, and \"Eating an apple\" includes \"Pick up the remote,\" which conflicts with the task description. Please normalize the scenarios and ensure the instructions match the intended task.","section":"Table 1"},{"comment":"The SUS result is reported only as M=83.21, SD=7.34; with N=12, a confidence interval or boxplot of individual scores would strengthen the usability claim.","section":"§6.2.3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid AR authoring system contribution, worth engaging with, but the context-awareness claim is softer than the abstract suggests. The genuinely new part is the integration: ChatGPT for step generation, user navigation and screenshots for context, per-frame text-conditioned motion diffusion (modified from MDM/GMD), and sigmoid-blend temporal smoothing. The per-frame conditioning change is real, and the temporal smoothing is simple and seems effective at reducing transition distance.\n\nWhat the paper does well: it identifies a real gap—AIGC lacks spatial/temporal context for AR instruction—and proposes a workflow that lets authors provide context by walking around. The two user studies are reasonable for a systems paper, and the SUS score (83) and preference over the PbD baseline are decent evidence that the authoring experience is usable. The authors are honest about the hand-object interaction limitation and the rigid-object scope.\n\nSoft spots, in proportion: First, the quantitative evaluation of the context-awareness claim rests on an unmeasured perception stack. The YOLO detector and MegaPose 6DoF on HoloLens screenshots have no accuracy or recall numbers in the study environments. The authors provide a manual 6DoF adjustment step, so failed perception doesn't break authoring, but it means we can't tell whether the observed spatial alignment came from perception or manual fixes. The 0.1m spatial distance claim is a single aggregate in Table 2 without per-scenario variance; the authors even note the Exercising task has the highest error. Second, User Study 2 runs many paired tests without multiple-comparison correction; a few p-values are in the 0.03-0.05 range, so some could be false positives. Third, the PbD baseline is home-built and likely less mature than state-of-the-art, which may inflate the comparison. These are real but not fatal; they are typical of CHI systems papers and mostly fixable with more transparent reporting.\n\nWho it's for: researchers in AR/HCI working on authoring tools or AIGC for AR. It deserves a serious referee; the system is novel enough and the user studies meaningful enough to warrant review, though heavy revision on the evaluation reporting is needed. I'd send it to review but tell the authors to report perception accuracy, add error bars, and correct for multiple comparisons.","headline":"A useful AR authoring pipeline with a real diffusion-model tweak, but the spatial context claim needs better perception evidence and more careful statistics.","tokens_in":31932,"tokens_out":1927,"would_cite":true,"duration_ms":19727,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By walking through a space, authors can give generative AI the spatial and temporal context it needs to make avatar-based AR instructions, with a diffusion-model change and temporal smoothing cutting motion discontinuity to 0.03 m while…","keywords":["augmented reality","generative AI","humanoid avatar instructions","context-aware authoring","motion diffusion model","temporal smoothing","text-to-motion","AR instruction authoring"],"falsifier":"Run CARING-AI in the same kitchen and living-room settings with ground-truth markers at each object, and compare the generated avatar's hand positions against those markers at interaction frames; if the root-to-object distance exceeds 0.1 m or the transition discontinuity stays near the 0.15 m baseline, the spatial and temporal context claims fail.","tokens_in":30986,"feed_emoji":"🥽","tokens_out":8321,"duration_ms":77659,"temperature":0.7,"pith_summary":"This paper tries to establish that the missing ingredient for AI-generated AR instructions is context, and that a short authoring walk can supply that context. CARING-AI lets a user speak a task, edit the resulting step list, and then walk through the environment while the headset records the trajectory and object locations; a modified diffusion model turns the text plus this context into humanoid-avatar animation overlaid in AR. The quantitative claim is that the modification plus temporal smoothing cuts motion discontinuity between steps from 0.15 m to 0.03 m while keeping spatial alignment under 0.1 m. In two user studies, participants rated the system usable and found it less mentally and physically demanding, faster, and less error-prone than authoring by physical demonstration. If the claims hold, non-experts can author situated AR instructions without programming, motion-capture hardware, or performing the task themselves.","feed_headline":"Walk the room, say the task: AR how-tos appear","feed_subtitle":"Smoothing cuts generation jumps to 0.03 m while avatars stay aligned with real objects, no code or mocap required.","key_machinery":"The carrying engine is a per-frame-conditioned motion diffusion model: instead of conditioning the entire motion sequence with one text embedding, each frame receives its own action and trajectory condition, so multiple actions can be generated in one pass. Around this sits a temporal smoothing function that blends the last frames of one action with the first frames of the next using a shifted sigmoid weight $f(K^1_t,K^2_t,\\alpha_t)=\\alpha_t K^1_t+(1-\\alpha_t)K^2_t$, then linearly resamples the blended segment back to full length. The scan module supplies the context: the author's walked trajectory gives global spatial routing, and object detection plus six-degree-of-freedom pose estimation on headset screenshots gives local object anchors for hand-object interactions. A large language model refines spoken task descriptions into editable step labels drawn from the motion dataset's vocabulary, keeping the generated text compatible with the motion generator.","core_discovery":"The paper argues that state-of-the-art text-to-motion generation is context-blind: it produces motions that are spatially offset from objects, temporally disconnected between steps, and presented at the wrong scale for AR instruction. CARING-AI's contribution is a workflow that feeds context into the generator by having the author walk through the target space and take screenshots at interaction locations. The system then generates humanoid-avatar animations that are globally routed along the author's trajectory, locally grounded at detected objects through six-degree-of-freedom pose estimation, and temporally stitched by a smoothing function so multi-step instructions play as one continuous demonstration. The reported result is a transition distance of 0.03 m versus 0.15 m for the baseline, with avatar-to-object absolute distance staying under 0.1 m.","pith_inferences":["If the workflow is sound, the same walk-to-contextualize loop should transfer to non-avatar AR cues such as arrows, text labels, images, and video, since those cues also need spatial placement and timing; the paper names this direction but does not build it.","Because context is captured separately from the motion generator, a CARING-AI-style pipeline could regenerate instructions for a new room by re-scanning, without rewriting the task description.","The per-frame conditioning plus sigmoid blending is a general recipe for stitching diffusion-generated motion clips into arbitrarily long sequences, and could be tested on motion vocabularies beyond the one used here.","The remote-authoring scenario suggests the headset scan itself could be replaced by any aligned spatial map, letting an author create instructions in one environment and deploy them in another."],"forward_implications":["Non-experts can author animated AR how-tos from speech plus a short walkthrough, with no programming and no motion-capture hardware.","Multi-step instructions render as one continuous animation instead of independent clips, removing the visible breaks that made earlier AI-generated tutorials feel disjointed.","The same textual instruction can be recontextualized in a new physical space by re-scanning, so instructions adapt to changed room layouts.","Compared with demonstration-based authoring, users in the study experienced lower mental and physical demand, fewer errors, and faster authoring time.","Spatial grounding keeps the avatar within 0.1 m of target objects, meeting the threshold the authors cite for plausible motion."],"supporting_citations":[{"why":"Supplies the guided diffusion baseline with trajectory guidance that the paper extends and compares against.","marker":"[55]"},{"why":"Base motion diffusion architecture whose per-frame conditioning is modified to generate multi-action sequences.","marker":"[118]"},{"why":"Dataset providing the action-label vocabulary and motion training data used by the generator and baseline.","marker":"[39]"},{"why":"Text-guided hand-object interaction motion generation model used for local interaction clips.","marker":"[15]"},{"why":"Supplies object six-degree-of-freedom pose estimation for overlaying virtual objects and grounding local interactions.","marker":"[62]"},{"why":"Object detector that supplies semantic object classes and locations from headset screenshots.","marker":"[101]"},{"why":"Large language model API that refines spoken tasks into editable step-by-step instructions aligned to the motion vocabulary.","marker":"[88]"},{"why":"Defines the transition-distance metric used to quantify temporal discontinuity between generated motions.","marker":"[3]"}],"fun_headline_variants":["Walk the room, say the task: AR how-tos appear","Context-aware AR authoring: walk to generate avatars","AI-generated AR avatars grounded in your walk","From walking to AR instructions: AI does the work","0.03 m precision: walking makes AR avatars stick"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The local-spatial grounding relies on the headset's camera accurately detecting each relevant object and locating it in six degrees of freedom from a screenshot; if an object is missed or its location is wrong, the avatar interacts with empty space and the context-awareness claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Walk the room, say the task: AR how-tos appear","Context-aware AR authoring: walk to generate avatars","AI-generated AR avatars grounded in your walk","From walking to AR instructions: AI does the work","0.03 m precision: walking makes AR avatars stick"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1686,"prompt_tokens":928,"completion_tokens":758,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":677}},"tokens_in":544,"tokens_out":758,"duration_ms":8068,"temperature":1.0,"reasoning_tokens":677,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T12:19:54.249300+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CARING-AI in the same kitchen and living-room settings with ground-truth markers at each object, and compare the generated avatar's hand positions against those markers at interaction frames; if the root-to-object distance exceeds 0.1 m or the transition discontinuity stays near the 0.15 m baseline, the spatial and temporal context claims fail.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Object detector that supplies semantic object classes and locations from headset screenshots."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Large language model API that refines spoken tasks into editable step-by-step instructions aligned to the motion vocabulary."}],"review_version":1}