{"id":"817f217a-77b2-4689-bbd4-b37ec3c0f7e0","arxiv_id":"2505.17389","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Breaking long manipulation tasks into atomic subtasks and collecting demonstrations from varied starting poses improves imitation learning success using fewer demonstration frames.","lead":"HD-Space is a data collection scheme that breaks long robot tasks into small steps, such as opening a box or picking up a cup, and records a short demonstration for each step from many different starting positions. In simulated and real-world tests, policies trained on these short demonstrations succeeded more often, with fewer recorded frames, than policies trained on full-task demonstrations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HD-Space's long-horizon claim is unsupported without a pure-HD control and boundary-state analysis; all real-world ablations mix HD atomic segments with naive full-task demos, so the composition mechanism is never isolated.","rationale":"The reader's weakest-assumption analysis points at the unvalidated chaining of atomic subtasks, and I agree that this is the core risk. My stress-test sharpens it: the experimental design cannot separate HD-Space's contribution from the effect of adding short, focused clips to a base of 25 naive full-task demos. Because HD-Space episodes are atomic segments with far fewer frames, equal episode counts do not mean equal task coverage, and the absence of any H-only condition leaves the central \"HD-Space-based data\" claim underdetermined. The long-horizon claim further depends on boundary generalization, which is asserted but never measured. The paper does provide consistent mixed-data gains, qualitative failure analysis, and a clear mechanism, so the evidence is not negative; it is incomplete in a way that a conditional verdict captures. Since the reader already assigned CONDITIONAL, my concern does not move the verdict, hence UNCHANGED.","tokens_in":11274,"tokens_out":5914,"duration_ms":77566,"concrete_test":"Run a three-arm ablation on the \"put the teacup into the box\" task with the same ACT+DINOv2+MAE policy: (i) N50 full-task demonstrations only, (ii) H50 atomic-segment demonstrations only with balanced coverage across the four subtasks, and (iii) N25+H25 as in the paper, matching total frames as closely as possible. Evaluate all three with 50 episodes x 3 seeds. If H50 does not reach or exceed N50 task-level success, HD-Space alone does not support the long-horizon claim; if it does, additionally record the end-effector pose at each atomic boundary during rollouts and compare it against the start-pose distribution used to collect the next atomic segment, reporting the fraction of boundary states that fall inside that distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 asserts that overlapping boundaries \"ensure the smooth execution of extended manipulation sequences\", but no experiment directly tests this chaining mechanism. In Fig. 5 and Table 2, every HD-Space condition is a mixture (N25/H25, N25/H50, N25/H75) that always includes 25 naive full-task episodes, while Table 3 shows HD-Space episodes are short atomic segments (e.g., 120 frames vs. 1300 naive frames for the electronic-pen task). An \"episode\" is therefore not a comparable unit across the two methods: equal episode counts imply very different full-task coverage. The observed gains could come from adding short, relatively uniform clips to a base of full-task demos rather than from the hierarchical atomic-space mechanism. The central long-horizon claim specifically requires that the end state of each atomic rollout falls inside the start-pose distribution of the next atomic space; no rollout-level measurement of boundary coverage is provided, so subtask-to-subtask error accumulation is unquantified. Without a pure-HD condition and boundary-overlap statistics, the causal claim that HD-Space data alone enables longer-horizon tasks with fewer demonstrations is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HD-Space, a data collection scheme for imitation learning in long-horizon manipulation. The method decomposes a full task into atomic subtasks, collects human demonstrations that start from varied end-effector poses around each subtask target, and then trains a single policy on a mixture of naive full-task demonstrations and these atomic segments. The authors evaluate the scheme with ACT-style policies on real-world tasks (teacup-in-box, mobile conveyor bowl/spoon grasping, unordered electronic-pen grasping), reporting higher success rates and more completed subtasks under equal episode counts, as well as fewer frames per episode. Two simulation tasks are used only to validate a stronger backbone (ACT+DINOv2+MAE), not the HD-Space mechanism itself.","tokens_in":11563,"tokens_out":6068,"duration_ms":49492,"significance":"If validated, HD-Space would be a simple, practical data-collection protocol that reduces demonstration cost and improves robustness for long-horizon imitation learning. The paper's strengths include real-robot experiments across several task families, a concrete hyperparameter appendix, and a public project page with videos. However, the current evidence does not isolate the proposed mechanism: all comparisons mix full-task demonstrations with atomic segments, episode units are not comparable because HD-Space segments are much shorter, and the claimed composition mechanism is never directly measured. The abstract's claim of evaluation across two simulated and five real-world tasks also overstates what is presented.","major_comments":[{"comment":"All HD-Space conditions contain 25 naive full-task episodes (N25/H25, N25/H50, N25/H75), so the comparison is between full demonstrations only and full demonstrations plus atomic segments. This does not isolate the hierarchical-atomic-space mechanism; the observed gains could come from adding short, varied clips to a base of full-task demonstrations. The central claim that HD-Space alone enables longer-horizon tasks with fewer demonstrations requires a pure-HD condition (e.g., H50, H75, or H100 with no naive episodes) or an appropriate control matched for data content.","section":"Section 4.2, Figure 5 and Table 2"},{"comment":"Episode counts are not comparable across methods: for the electronic-pen task, a naive episode has 1300 frames while an HD-Space episode has 120 frames. Equal episode counts therefore imply substantially fewer total frames for the HD-Space conditions, and 'fewer episodes' conflates episode length with data volume. The success-rate comparisons should be reported against matched total frame counts or otherwise analyzed per frame.","section":"Section 4.2, Table 3 vs Figure 5"},{"comment":"The claim that overlapping atomic-space boundaries 'ensure the smooth execution of extended manipulation sequences' is asserted but not verified. No rollout-level statistic shows that the terminal states of atomic segments fall inside the start-pose distribution of the following segment, nor that errors do not accumulate at segment transitions. The paper should provide boundary-overlap coverage statistics or an end-to-end rollout analysis of how atomic chunks compose.","section":"Section 3, 'overlapping boundaries'"},{"comment":"The simulation experiments evaluate only the model backbone (ACT vs ACT+DINOv2+MAE) and do not evaluate HD-Space itself. The abstract's statement that the method was evaluated 'across two simulated and five real-world long-horizon manipulation tasks' is therefore misleading. In addition, Ref. [50] (ADC) is described as the closest prior method, but no comparison to ADC is run, and the paper gives no justification for omitting it as a baseline.","section":"Section 4.1, Table 1, and Abstract"},{"comment":"Success rates and completed-subtask counts are reported without error bars, standard deviations, or trial counts. Given that some reported differences are as small as 8 percentage points, the results need variance information and, ideally, statistical tests across seeds or evaluation trials before the claimed improvements can be assessed.","section":"Figures 5 and Table 2"}],"minor_comments":[{"comment":"The abstract says five real-world tasks, while Section 4.1 says three; clarify the count, for example by treating the mobile conveyor task with its speed conditions as separate task variants.","section":"Abstract and Section 4.1"},{"comment":"The phrase 'segments tdivides' appears to be a typo; it should read 'segments' or 'divides'.","section":"Section 3, first paragraph"},{"comment":"The notation for the action sequence horizon N and the subscript on the expectation is implicit; define these explicitly.","section":"Section 3, Eq. (1)"},{"comment":"The protocol for choosing atomic task boundaries and for 'evenly placing' start poses around the target is underspecified; report the sampling radius, density, and any validation that the segmentation captures decision-relevant structure.","section":"Section 3, atomic segmentation and sampling"},{"comment":"The heading 'Max lenght' should be 'Max length'.","section":"Table 2"},{"comment":"The claim that Ref. [62] supports the conveyor-belt speed challenge is unclear, since [62] appears to be a streaming-perception paper rather than a manipulation source; please clarify or replace the reference.","section":"Section 4.1, conveyor task"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a plausible data-collection contribution, and the real-robot experiments are a strength. The main scientific risk is overclaiming relative to the current evidence, particularly the missing pure-HD control and the lack of direct measurement of the composition mechanism. I believe these gaps can be addressed with additional experiments within the paper's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: HD-Space is a genuinely useful idea for cutting IL data cost, but the experiments don't isolate the mechanism they claim. The core idea—segment a long-horizon task into atomic subtasks and collect human demos from varied start poses around each target—is simple, plausible, and new as far as I know. The consistent gains across several real and simulated tasks suggest something real is going on.\n\nThe paper does a lot right: it targets a practical pain point (data cost in IL), the method is easy to describe and could transfer to other policy classes, and the qualitative failure analysis on the spoon task gives some intuition for why varied start poses help.\n\nThe soft spot is in the comparison. In every real-world ablation, the HD-Space condition is a mixture: N25/H25, N25/H50, N25/H75—always with 25 naive full-task episodes. There is no pure-HD condition. So the gains could just come from adding short, relatively uniform clips to a base of full demos, not from the hierarchical atomic-space mechanism. That is the stress-test's main point, and it holds up. Table 3 makes it worse: HD-Space episodes are 100–400 frames versus 350–1300 for naive, so 'same episode count' is not the same data volume. The long-horizon claim specifically requires that the end of each atomic rollout falls inside the next atomic start distribution; the paper never measures boundary overlap or error accumulation across subtask transitions.\n\nMinor: abstract says five real tasks, body says three; no error bars or trial counts; no comparison against ADC, which is the closest prior method and is cited but never run.\n\nOverall, the central idea is good and the empirical pattern is encouraging, but the causal story is not yet established. I'd send it to review, but the referee should require a pure-HD condition, a comparable-data-volume comparison, and an ADC baseline.","headline":"HD-Space is a promising data-collection idea, but the experiments currently conflate adding short clips with the hierarchical mechanism.","tokens_in":12077,"tokens_out":2592,"would_cite":true,"duration_ms":30415,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T40","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Segmenting a robot task into atomic subtasks and collecting each one from varied start poses around the target—the HD-Space scheme—lets imitation learning policies train from smaller datasets while improving success rates on long-horizon…","keywords":["imitation learning","robotic manipulation","data collection space","long-horizon manipulation","hierarchical task segmentation","behavior cloning","action chunking transformer"],"falsifier":"Run a controlled comparison on a single long-horizon task where the object's initial pose is moved to positions that fall outside the region spanned by the evenly placed start poses, keeping all other collection details identical; the claim that HD-Space covers error-prone states predicts a large success-rate drop for these out-of-reach configurations, while the naive full-trajectory method should degrade similarly. A result showing equal performance would falsify the coverage mechanism.","tokens_in":11064,"feed_emoji":"🤖","tokens_out":7073,"duration_ms":49611,"temperature":0.7,"pith_summary":"The paper proposes HD-Space, a way of collecting human demonstrations for imitation learning that splits a long-horizon manipulation task into a sequence of atomic subtasks and records each subtask separately, starting from several hand-chosen poses around the target. The claim is that these atomic start poses expose the policy to the states that actually accumulate errors during execution, so the trained model becomes more robust without needing more demonstrations. If this is right, the cost bottleneck of imitation learning is partly a data-collection problem, not just a model-size or data-volume problem: smaller, smarter datasets can train policies that complete longer task sequences. The paper supports the claim with experiments on two simulated and five real-world long-horizon tasks, comparing HD-Space data against the naive full-trajectory collection method at matched episode counts.","feed_headline":"Subtask-split demos beat full-trajectory demos on real robot tasks","feed_subtitle":"Starting each subtask from varied poses covers error-prone states, so smaller datasets train better long-horizon policies.","key_machinery":"The central object is the atomic data collection space: for each hand-selected atomic subtask, the robot's end-effector is placed at a set of start poses evenly distributed around the target point, and human demonstration is recorded from those poses until the atomic action ends. The overlapping boundaries between consecutive atomic spaces are what let the individual chunks chain into a full long-horizon execution. This design does the work of turning a single narrow trajectory distribution into a broader, proactive coverage of the states where compounding errors typically happen.","core_discovery":"The central discovery is that where you press 'record' in human demonstration matters as much as how many times you record. HD-Space segments a fine manipulation task into semantically meaningful atomic tasks (e.g., open box, pick up cup, put in box, close box). For each atomic task, the human demonstrator starts the robot end-effector at several evenly placed poses around the target—rather than at one fixed start pose—and records until the atomic action completes. Consecutive atomic spaces overlap at their boundaries, which the paper argues ensures that long-horizon sequences chain smoothly. This yields a dataset that covers more of the state/action space that is error-prone during execution, so the trained policy 'sees as many states around the target point as possible' and suppresses erroneous predicted trajectories. Across their evaluations, HD-Space data improves success rates, increases the mean number of completed sequential subtasks, and reduces the number of frames recorded per episode, making the method both more robust and cheaper than naive collection.","pith_inferences":["If the coverage mechanism is the real driver, HD-Space is effectively a state-space prior; the natural next step is to derive the sampling radius and density of start poses from the task geometry or from the current policy's failure modes rather than hand-tuning them.","A testable extension is to measure the distribution of policy confidence (or action variance) at the overlapping boundary states; the paper's chaining argument predicts these boundaries are the most failure-prone points and should receive proportionally more demonstrations.","HD-Space could act as a cheap bootstrap for human-in-the-loop and simulation-based pipelines, since a stronger baseline shortens the error-waiting phase; the paper notes orthogonality to these approaches but does not test the combination.","The method's benefit seems to scale with task difficulty: the largest gains appear in tasks requiring precise angle, height, and speed matching, suggesting a quantitative relation could be established between the size of the error-prone region and the number of start poses needed."],"forward_implications":["At matched training episode counts, HD-Space data improves success rate over naive full-trajectory data by 8 percentage points on the teacup-into-box task, by 18–32 points on mobile bowl grasping, and by 20–44 points on spoon grasping.","On mobile grasping tasks, 25 naive plus 25 HD-Space episodes outperform 100 naive episodes, indicating the method extracts more useful signal per episode.","HD-Space reduces the number of frames recorded per demonstration (e.g., from 600 to 400 for the teacup task and from 1300 to 120 for grabbing electronic pens), which lowers human demonstration cost directly.","The average number of completed sequential subtasks increases when HD-Space data is added, so the trained policy extends to longer-horizon execution.","Because the scheme is a data-collection change rather than a network change, the gains are expected to combine with stronger policy architectures such as those using richer visual features."],"supporting_citations":[{"why":"the action-chunking transformer policy used for all training and the simulated benchmarks it is compared against","marker":"[5]"},{"why":"documents the compounding-error problem in trajectory-based imitation learning that motivates covering states around waypoints","marker":"[28]"},{"why":"the closest alternative active data-collection method, using adversarial human-robot interventions, which HD-Space is compared with conceptually","marker":"[50]"},{"why":"human-in-the-loop remote teleoperation, the passive-correction approach HD-Space is designed to replace","marker":"[16]"},{"why":"the self-supervised vision backbone that carries the stronger model used in real experiments","marker":"[57]"},{"why":"the masked-autoencoder augmentation applied in all real-world model training","marker":"[60]"},{"why":"the low-cost teleoperation robot platform used for the desktop real-world task","marker":"[10]"}],"fun_headline_variants":["Subtask pose variety boosts robot learning with fewer demos","More start poses, less demo data for long-horizon robot skills","Robots learn better from subtask demos with varied starting poses","HD-Space: smarter recording spots train stronger robot policies","Pose diversity in subtasks slashes demo count for robot tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method relies on a human correctly choosing the atomic task boundaries and the spread of start poses around each target; if those choices miss states that matter, the collected data will not cover the genuinely error-prone region, and the supposed advantage over full-trajectory collection collapses.","fun_headline_variants_meta":{"raw":{"variants":["Subtask pose variety boosts robot learning with fewer demos","More start poses, less demo data for long-horizon robot skills","Robots learn better from subtask demos with varied starting poses","HD-Space: smarter recording spots train stronger robot policies","Pose diversity in subtasks slashes demo count for robot tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":3048,"prompt_tokens":992,"completion_tokens":2056,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1968}},"tokens_in":608,"tokens_out":2056,"duration_ms":11310,"temperature":1.0,"reasoning_tokens":1968,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:47:34.971109+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison on a single long-horizon task where the object's initial pose is moved to positions that fall outside the region spanned by the evenly placed start poses, keeping all other collection details identical; the claim that HD-Space covers error-prone states predicts a large success-rate drop for these out-of-reach configurations, while the naive full-trajectory method should degrade similarly. A result showing equal performance would falsify the coverage mechanism.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"documents the compounding-error problem in trajectory-based imitation learning that motivates covering states around waypoints"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the masked-autoencoder augmentation applied in all real-world model training"}],"review_version":1}