{"id":"bd430da5-6f0b-4314-82a1-cb97f32fcb15","arxiv_id":"2505.09430","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Mini Diffuser trains multi-task robotic diffusion policies by reusing each vision-language condition for many noised action samples, reaching 77.6% RLBench success with 4.8% of the training time and 6.6% of the memory of 3D Diffuser Actor.","lead":"This paper introduces a training trick for robotic diffusion policies: pair several noisy action samples with the same camera and language condition, instead of one sample per condition. The result is a policy that keeps about 95% of the top method's success rate while cutting training time and memory by more than an order of magnitude.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline efficiency ratios are not controlled for training budget: Table 1 compares Mini-Diffuser on one 4090 with 3D Diffuser Actor on 6 A100s without reporting gradient steps or unique demonstration counts per step, so the 20x speedup may be dominated by training fewer updates rather than by…","rationale":"The paper's central claim is an efficiency claim, and the most load-bearing condition is that the 4.8% time and 6.6% memory ratios reflect the method's contribution rather than uncontrolled differences in training budget, step count, and hardware. Table 3 empirically supports the scaling premise in the tested regime, so I do not see the reader's scaling concern as the primary threat: even if no general bound is given, the M=64 overhead is small for the action-token lengths used here. The unresolved issue is the denominator: the headline comparison mixes hardware, backbone changes, and the number of gradient updates. The paper deserves credit for honest ablations, the 'fake batch' admission in Sec. 5.2.1, and real-world validation; the concern is about comparability, not algorithmic correctness. If the released training logs can confirm that Mini-Diffuser sees a comparable number of unique demonstrations and still reaches 95% of 3D Diffuser Actor in ~20x less wall-clock, the claim would stand; if the speedup depends mainly on training fewer steps, the paper should say so explicitly and decompose the savings. This is why CONDITIONAL remains the appropriate verdict, and no shift to ACCEPT or REJECT is warranted without the training-budget check.","tokens_in":13586,"tokens_out":18257,"duration_ms":185712,"concrete_test":"Use the released training logs to count, for the Table 1 Mini-Diffuser run, the total number of optimizer steps and the per-step (B, M) schedule; obtain or reconstruct the corresponding 3D Diffuser Actor run. Then, on a single RTX 4090 with matched optimizer settings and equal total number of unique demonstrations seen, measure wall-clock for both methods to reach the success levels of Table 2. If the recomputed Mini-Diffuser time ratio is close to the advertised 4.8%, the central claim holds; if the gap largely disappears, the 20x claim is dominated by training budget and hardware, not by Level-2 batching.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The advertised efficiency ratios (Sec. 5.1.3, Table 1: 95.4% of 3D Diffuser Actor success at 4.8% time and 6.6% memory) compare Mini-Diffuser on a single RTX 4090 for 13 hours with 3D Diffuser Actor on 6 A100s for 6 days. The paper does not state how many gradient updates each run performed, nor how many unique (state, action) demonstrations were seen per update. Table 3 shows that per-iteration cost at B=100, M=64 is only 106.3% of B=100, M=1, and reverting the PTv3 backbone increases memory/time by roughly 18-44%; these per-step factors cannot explain a ~20x wall-clock gap on one GPU versus six. Unless Mini-Diffuser also converges in many fewer updates, the 20x claim is not explained by the proposed mechanism. The text acknowledges in Sec. 5.2.1 that M>1 samples reuse the same condition and are 'fake' batches, so the independent-demonstration count per step is B, not B*M. Without reporting B, M, and total steps for the Table 1 run, the headline time ratio is not a controlled measure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Mini-Diffuser, a diffusion policy for multi-task vision-language manipulation that pairs B state-action conditions with M noised action samples per condition, a scheme the authors call two-level minibatching. To make this efficient, the architecture uses masked global attention to prevent inter-sample leakage, a local kernel-based query for 3D grounding, and per-sample FiLM modulation, with a PTv3-style U-Net backbone and a decoupled action head. On RLBench, the authors report 77.6% average success (95.4% of 3D Diffuser Actor) with normalized training time 1.9 vs 39 and memory 16 vs 240 GB, and they include real-world success rates on twelve tasks. Ablations vary the Level-2 batch size M and remove components such as 3D RoPE, the PTv3 backbone, and the local convolution.","tokens_in":13812,"tokens_out":6089,"duration_ms":59998,"significance":"If the efficiency claims hold, the result is practically significant: it suggests near-state-of-the-art multi-task manipulation policies can be trained on a single consumer GPU in about 13 hours, which would substantially lower the barrier to reproducing and extending diffusion-policy methods. The paper's strengths include evaluation on an external benchmark, real-world experiments, an ablation that isolates the Level-2 batching contribution, and Fig. 4, which shows faster convergence per gradient step as M increases. However, the headline time and memory ratios in Table 1 are not controlled comparisons: the memory column appears to aggregate across multiple GPUs for baselines, and the time comparison does not report total gradient steps or epochs. The central efficiency claim is plausible but needs to be substantiated with step counts, epochs, and per-device memory before it can be accepted at face value.","major_comments":[{"comment":"The headline efficiency ratios (4.8% time, 6.6% memory) are not controlled comparisons. The memory column appears to be aggregate across multiple GPUs: PerAct trained on 8 V100s reports 128 GB (16 GB per GPU), and 3D Diffuser Actor trained on 6 A100s reports 240 GB, while Mini-Diffuser reports 16 GB on a single RTX 4090. Comparing aggregate multi-GPU memory with single-GPU memory conflates algorithmic memory savings with hardware scaling. Please report per-device peak memory and total aggregate memory for every method, and state explicitly whether the normalized time unit already accounts for GPU count and type.","section":"Table 1; §5.1.3"},{"comment":"The 4.8% time claim is not controlled for training budget. The paper does not report the total number of gradient steps or epochs for the Mini-Diffuser run in Table 1, nor the corresponding numbers for 3D Diffuser Actor. Since Table 3 shows only 106.3% per-iteration time for B=100, M=64 relative to B=100, M=1, the roughly 20x reduction in normalized time must arise primarily from requiring far fewer optimizer steps; Fig. 4 suggests this, but the actual step counts, data repeats, and learning-rate schedules are missing. Please provide a same-hardware or same-step-budget comparison, or at minimum report the total steps, epochs, and unique demonstrations for every run in Table 1.","section":"§5.1.3; §5.2.1"},{"comment":"The statement that M=64 processes 64 times more training samples per step is not equivalent to seeing 64 times more independent demonstrations, because Level-2 samples reuse the same condition and the paper itself calls them 'fake' batches. Please report the number of unique (state, action) pairs processed per step for each configuration, and report wall-clock time to reach a fixed success threshold for M=1, 4, 16, 64, 128, and 256, so that the trade-off between per-step cost and convergence speed can be evaluated directly.","section":"§5.2.1; Eq. (10)"}],"minor_comments":[{"comment":"The performance ratio appears as 95.4% in Table 1, 95.6% in the text of §5.1.3, and roughly 95% in the abstract; these numbers should be made consistent.","section":"Abstract; Table 1; §5.1.3"},{"comment":"There are several typos in this section, including 'foundition', 'spaitial', and 'set if tasks'; a careful proofreading pass is needed.","section":"§2.1"},{"comment":"The caption says 'Duo-Level batches' instead of 'Two-Level batches', and the column header 'per decent' should be 'per descent'; the meaning of the first column header 'consistent' is unclear and should be defined.","section":"Table 3"},{"comment":"The caption contains the typo 'baed' and the bubble percentages duplicate numbers in the legend; please clarify what is being plotted.","section":"Figure 2 caption"},{"comment":"Real-world results are based on only 10 trials per task and no error bars are reported; the table caption should state this limitation explicitly.","section":"Table 4"},{"comment":"The introduction claims the architecture adds 'no additional memory usage or computational overhead,' but Table 3 reports 102.2% memory and 106.3% time at B=100, M=64 relative to M=1; please rephrase to 'near-zero' or 'a few percent' overhead.","section":"§1; §5.2 (Table 3)"}],"recommendation":"major_revision","confidential_remarks":"The central contribution is attractive and the internal ablation supports the mechanism, but the headline efficiency ratios in Table 1 are the main selling point and they are not yet a controlled measurement. I would require, at minimum, explicit reporting of gradient steps, epochs, unique demonstrations, and per-device memory, and ideally a same-hardware comparison or a same-step-budget comparison, before the paper can be accepted. The code and training logs are promised on a website; the reviewer did not have access to them, so the reproducibility claims could not be verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth sending out for review. The core idea is simple and mostly works: in action diffusion the condition is big and the target is small, so you can pair many noised action samples with one shared condition and get most of the benefit of a large batch at a fraction of the cost. The two-level batching strategy and the masked-attention/FiLM architecture that make it leak-free are the genuinely new pieces; the components are known, but the combination is not in the cited prior work.\n\nThe empirical core is honest. Table 3 is the best part: B=100, M=64 costs 102.2% memory and 106.3% time per step versus B=100, M=1, and it lifts success from 44.1% to 78.3% after the same number of steps. That is a real sample-efficiency gain, and Fig. 4 shows the convergence benefit directly. The ablation also correctly does not attribute the gains to pretrained PTv3 weights.\n\nThe soft spots are mostly about the headline comparisons. The 4.8% training-time claim in Table 1 compares a single 4090 for 13 hours against 6 A100s for 6 days without reporting how many gradient updates or unique demonstrations each run saw. The per-step numbers alone (106%) cannot explain a ~20x wall-clock gap; the rest must come from the model needing fewer updates, which is exactly what Level-2 batching is supposed to deliver, but it is not stated for the main run. The stress-test note is right that the ratio is uncontrolled, but wrong to imply this undermines the mechanism—faster convergence per step is the mechanism, and Fig. 4 supports it. Reporting steps and a compute-matched baseline would fix the claim. The intro also overstates the overhead as 'zero' when Table 3 shows 3% memory and 7% time; a minor wording issue. There are no aggregate error bars on the average success rate, and the memory comparison mixes hardware, but neither changes the direction of the result.\n\nBottom line: this is a practical, believable efficiency method, with an honest ablation table and external benchmarks. It deserves a serious referee, though the revisions should force the authors to report gradient steps and ideally compare on matched hardware. Would bring it to a reading group.","headline":"A practical, believable efficiency method for training multi-task diffusion policies on a single GPU; the headline time ratio is not fully controlled, but the core mechanism checks out.","tokens_in":14392,"tokens_out":3341,"would_cite":true,"duration_ms":31174,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Mini Diffuser trains multi-task diffusion robot policies at about 5 percent of the usual cost, while keeping roughly 95 percent of the performance.","keywords":["multi-task robotic manipulation","diffusion policy","training efficiency","two-level mini-batching","RLBench","vision-language-action","transformer attention masking","imitation learning"],"falsifier":"A direct test is to fix the Level-1 batch and increase the number of action tokens per sample (for example, predicting longer action horizons) while holding M fixed; if per-step time and memory scale with the total action-token count instead of staying near the condition-token cost, the core efficiency claim fails.","tokens_in":1669,"feed_emoji":"🤖","tokens_out":2001,"duration_ms":66962,"temperature":0.7,"pith_summary":"This paper argues that diffusion-based robotic policies are expensive to train for the wrong reason: unlike image diffusion, the action output is low-dimensional while the conditioning input (visual scene, language, proprioception) is high-dimensional. Mini Diffuser exploits this asymmetry with two-level mini-batching, pairing many noised action samples with each shared condition instead of the usual one-to-one sampling. On the 18-task RLBench benchmark it reports 95.4 percent of the average success of 3D Diffuser Actor while using 4.8 percent of the training time and 6.6 percent of the memory, allowing end-to-end training on a single RTX 4090 in about 13 hours. If correct, this removes a major practical barrier to training multi-task diffusion policies, making them feasible for labs without multi-GPU clusters.","feed_headline":"Train diffusion robot policies on one GPU in 13 hours","feed_subtitle":"Mini Diffuser keeps 95% of top method performance using 5% of the training time and 7% of the memory.","key_machinery":"The mechanism is two-level mini-batching formalized in equations (6) through (10): sample B state-action pairs, then for each pair sample M independent noise-timestep tuples, yielding a flattened batch of B times M noised actions that share condition tokens. The supporting architecture replaces standard transformer self-attention with a masked attention matrix M_QK in equation (13), where action tokens attend within their own sample and to shared visual and context tokens, while condition tokens do not attend to actions; per-sample FiLM modulation in equation (18) conditions each sample on its noise step, and a local 3D kernel-based query in equation (17) gathers spatial context around query positions. These non-invasive components keep samples independent while processing many samples per condition in one forward pass, which is what cuts per-sample compute and memory.","core_discovery":"The central claim is that in action diffusion the condition is much larger than the target, so re-encoding the same condition for every noised action sample wastes nearly all the training cost. Mini Diffuser organizes training into Level-1 batches of B state-action pairs and Level-2 batches of M independent noise-timestep samples per condition, then flattens them into one batch of B times M samples while encoding each condition only once. A masked global attention mechanism lets action tokens attend to their own tokens and to shared condition tokens, but never to other action samples, and condition tokens never attend to actions, preventing information leakage. The paper reports that with M equal to 64, it processes 64 times more samples per gradient step for roughly 3 percent extra memory and 7 percent extra time, and that the resulting policy reaches 77.6 percent average success on RLBench versus 81.3 percent for 3D Diffuser Actor, with real-world experiments showing preserved multimodal behavior.","pith_inferences":["Beyond the paper, the same two-level batching idea could accelerate other conditional generative models whose target space is much smaller than the condition space, such as long-context control or video-conditioned action prediction.","A testable extension is to measure how the speed-up degrades as the action token sequence length L grows; if action tokens start to dominate the batch, the condition-reuse premise fails and conventional batching should be restored.","Because Level-2 batches share conditions, they reduce gradient diversity; reinvesting the saved compute into more Level-1 conditions or more demonstrations may close more of the remaining 5 percent performance gap than pushing M to very large values.","The paper's architecture suggests a general recipe: when the target is low-dimensional, batch over target samples per condition and design the network with masks and per-sample modulation to prevent leakage across those samples."],"forward_implications":["Multi-task diffusion policies become trainable end-to-end on a consumer GPU in about 13 hours instead of requiring multi-GPU clusters for days.","Effective batch size can be scaled by increasing M without a proportional increase in memory or compute, accelerating convergence in gradient steps.","The efficiency gain is orthogonal to inference accelerations like DDIM or flow matching, so deployment speed can be improved independently of training cost.","The approach preserves diffusion's multimodal action generation, as shown in real-world trials where the same visual scene with different language instructions produces distinct valid actions.","The remaining small performance gap relative to state-of-the-art methods shrinks further on several individual tasks, where Mini Diffuser outperforms 3D Diffuser Actor."],"supporting_citations":[{"why":"3D Diffuser Actor, the state-of-the-art diffusion policy that Mini Diffuser compares against; supplies the baseline success rate, training time, memory, and the architecture whose self-attention is replaced.","marker":"[7]"},{"why":"PerAct, which defines the RLBench benchmark protocol, task split, keypose segmentation, 100 demonstrations per task, and evaluation over 300 unseen episodes per task.","marker":"[21]"},{"why":"RLBench, the simulator and learning environment that provides the multi-task manipulation tasks and demonstrations used in all simulation experiments.","marker":"[38]"},{"why":"Act3D, the source of 3D point cloud feature fields and 3D rotary positional encoding used in the transformer architecture.","marker":"[22]"},{"why":"RVT-2, which supports the decoupled action head design where rotation and gripper states are predicted with classification heads conditioned on the denoised position.","marker":"[24]"},{"why":"Point Transformer v3, whose U-Net-style backbone and downsampling point indices are used to build the Mini-Diffuser architecture.","marker":"[36]"},{"why":"SAM2Act, the strongest baseline on the benchmark that Mini Diffuser compares against in the overall cost-performance comparison.","marker":"[19]"},{"why":"FiLM, the feature-wise linear modulation mechanism used for per-sample conditioning on the diffusion noise step.","marker":"[35]"}],"fun_headline_variants":["Mini Diffuser: 95% performance with 5% training time","Train robot policies 20x faster with Mini Diffuser","One GPU, 13 hours: Mini Diffuser trains robot policies","Slash diffusion policy training time by 20x with Mini Diffuser","Mini Diffuser: 5% training time, 7% memory, 95% performance"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The speed-up depends on condition tokens dominating the compute and memory cost of a training batch, so adding more action samples per condition adds only a few percent overhead; if action sequences grow long enough to dominate the token count, the savings shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Mini Diffuser: 95% performance with 5% training time","Train robot policies 20x faster with Mini Diffuser","One GPU, 13 hours: Mini Diffuser trains robot policies","Slash diffusion policy training time by 20x with Mini Diffuser","Mini Diffuser: 5% training time, 7% memory, 95% performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000979,"raw_usage":{"total_tokens":4172,"prompt_tokens":974,"completion_tokens":3198,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":3101}},"tokens_in":590,"tokens_out":3198,"duration_ms":22915,"temperature":1.0,"reasoning_tokens":3101,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:32:46.470235+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to fix the Level-1 batch and increase the number of action tokens per sample (for example, predicting longer action horizons) while holding M fixed; if per-step time and memory scale with the total action-token count instead of staying near the condition-token cost, the core efficiency claim fails.","supporting_citations":[{"cited_title":"Rlbench: The robot learning benchmark & learning environment,","cited_arxiv_id":null,"evidence_quote":"RLBench, the simulator and learning environment that provides the multi-task manipulation tasks and demonstrations used in all simulation experiments."},{"cited_title":"Act3D: 3D feature field transformers for multi-task robotic manipulation,","cited_arxiv_id":null,"evidence_quote":"Act3D, the source of 3D point cloud feature fields and 3D rotary positional encoding used in the transformer architecture."},{"cited_title":"RVT-2: Learning precise manipulation from few demonstrations,","cited_arxiv_id":null,"evidence_quote":"RVT-2, which supports the decoupled action head design where rotation and gripper states are predicted with classification heads conditioned on the denoised position."},{"cited_title":"SAM2Act: Integrating visual foundation model with a memory architecture for robotic manipulation,","cited_arxiv_id":null,"evidence_quote":"SAM2Act, the strongest baseline on the benchmark that Mini Diffuser compares against in the overall cost-performance comparison."},{"cited_title":"Film: Visual reasoning with a general conditioning layer,","cited_arxiv_id":null,"evidence_quote":"FiLM, the feature-wise linear modulation mechanism used for per-sample conditioning on the diffusion noise step."}],"review_version":1}