{ "plotting_plan": [ { "figure_id": "fig_method_overview", "title": "iMF-AttnRes VLA Overview", "plot_type": "diagram", "data_source": "idea.md", "objective": "Paper Banana prompt placeholder for a conceptual pipeline diagram showing visual observations and robot state entering a VLA policy, iMF predicting average flow with one-to-few inference evaluations, AttnRes aggregating depth-wise residual features, and chunked robot actions being executed in RoboIMI.", "aspect_ratio": "16:9" }, { "figure_id": "fig_socket_reward_latency", "title": "Socket Insertion Reward-Latency Tradeoff", "plot_type": "plot", "data_source": "experimental_log.md", "objective": "Paper Banana prompt placeholder for a scatter/bar hybrid plot comparing socket-insert avg_reward against avg_inference_time_ms or inference FPS for iMF-AttnRes, diffusion-style VLA, ACT, and SmolVLA, emphasizing that iMF-AttnRes reaches higher reward than diffusion-style baselines while using roughly 14--16 ms inference latency instead of hundreds of milliseconds.", "aspect_ratio": "4:3" }, { "figure_id": "fig_sim_transfer_ablation", "title": "Sim Transfer Ablation Summary", "plot_type": "plot", "data_source": "experimental_log.md", "objective": "Paper Banana prompt placeholder for a grouped bar chart over sim_transfer ablations showing avg_reward and success_like episode count for native diffusion policy, best iMF-AttnRes, multi-token ResNet18 iMF, full-AttnRes vision, and horizon/execution variants.", "aspect_ratio": "4:3" }, { "figure_id": "fig_imf_attnres_components", "title": "iMF and AttnRes Component Mechanics", "plot_type": "diagram", "data_source": "idea.md", "objective": "Paper Banana prompt placeholder for a technical diagram contrasting classical flow matching instantaneous velocity prediction, Mean Flow average velocity prediction, iMF's corrected instantaneous-velocity training relation, and AttnRes softmax aggregation over preceding residual outputs.", "aspect_ratio": "16:9" } ], "intro_related_work_plan": { "introduction_strategy": { "hook_hypothesis": "Diffusion-style action generators can produce expressive robot policies, but their iterative denoising or ODE-style sampling creates a latency bottleneck for closed-loop robot control.", "problem_gap_hypothesis": "Existing fast robot policies often trade away action-distribution expressivity or require distillation, while existing one-step flow methods have not fully explored improved mean-flow training and depth-wise residual aggregation in VLA imitation learning.", "search_directions": [ "Find foundational and survey papers on imitation learning and visuomotor policy learning for robotic manipulation before 2026-04-15", "Find papers on diffusion policy and action diffusion for visuomotor robot control before 2026-04-15", "Find papers on vision-language-action robot policies, including RT-1, RT-2, OpenVLA, SmolVLA, and flow-based VLA models before 2026-04-15", "Find papers on flow matching, rectified flow, Mean Flow, and improved Mean Flow for one-step or few-step generation before 2026-04-15", "Find papers on residual connections, Transformers, PreNorm residual architectures, hyper-connections, and attention residual aggregation before 2026-04-15" ] }, "related_work_strategy": { "overview": "Organize related work around robot action generation, fast flow-based generation, and residual/attention architecture design. The review should explain why a one-to-few-step iMF policy with AttnRes is a plausible alternative to iterative diffusion action generation.", "subsections": [ { "subsection_title": "2.1 Diffusion and Flow Policies for Robot Action Generation", "methodology_cluster": "Diffusion Policy, action diffusion, flow matching, and generative action policies", "sota_investigation_mission": "Identify influential robot manipulation policies that use diffusion, flow matching, or iterative generative sampling, and summarize their sampling-step and latency implications.", "limitation_hypothesis": "Iterative sampling improves expressivity but makes high-frequency closed-loop control expensive, especially when policies require tens or hundreds of inference evaluations per action chunk.", "limitation_search_queries": [ "Diffusion Policy Visuomotor Policy Learning via Action Diffusion latency sampling steps", "robot manipulation action diffusion policy inference speed sampling steps", "flow matching robot policy action generation one-step few-step" ], "bridge_to_our_method": "iMF-AttnRes targets the same expressive action-generation setting but trains the policy to predict an average flow usable with one-to-few inference evaluations." }, { "subsection_title": "2.2 Vision-Language-Action Models and Action Chunking", "methodology_cluster": "VLA policies, action tokenization, action chunking, and transformer-based robot policies", "sota_investigation_mission": "Survey VLA and transformer imitation-learning systems that condition robot actions on language, images, and robot state, including models that improve deployment speed via chunking or compact action representations.", "limitation_hypothesis": "Compact VLA and chunking approaches can improve speed, but may not directly address the multi-step generative sampling cost of diffusion-style policies or the training instability of naive one-step flow policies.", "limitation_search_queries": [ "OpenVLA vision language action robot policy imitation learning", "RT-1 RT-2 robotics transformer vision language action control", "Action Chunking with Transformers ALOHA imitation learning", "SmolVLA efficient vision language action robotics" ], "bridge_to_our_method": "The proposed policy keeps the VLA imitation-learning framing while replacing slow iterative action sampling with iMF and strengthening deep policy representation with AttnRes." }, { "subsection_title": "2.3 Mean-Flow Training for One-Step Generation", "methodology_cluster": "Flow matching, rectified flow, Mean Flow, improved Mean Flow, and fast-forward generative models", "sota_investigation_mission": "Trace the development from flow matching and rectified flow to Mean Flow and improved Mean Flow, focusing on how average velocity prediction enables one-step or few-step generation.", "limitation_hypothesis": "Naively substituting conditional velocities into nonlinear mean-flow training relations can introduce unstable objectives; improved Mean Flow addresses this with a corrected training relation that better preserves flow-matching stability.", "limitation_search_queries": [ "Mean Flows for One-step Generative Modeling average velocity", "Improved Mean Flows fast-forward generative models conditional velocity marginal velocity", "Flow Matching for Generative Modeling conditional vector field", "Rectified Flow generative modeling one step sampling" ], "bridge_to_our_method": "We adapt the improved mean-flow idea from image generation to robot action generation, where lower sampling latency directly affects control frequency." }, { "subsection_title": "2.4 Residual Aggregation and Attention Residuals in Deep Transformers", "methodology_cluster": "Residual connections, PreNorm Transformers, hyper-connections, and AttnRes", "sota_investigation_mission": "Investigate why residual paths are central to deep networks and how learned or attention-based residual aggregation can mitigate feature dilution and gradient imbalance in deep transformers.", "limitation_hypothesis": "Standard unit-weight residual accumulation treats all previous layer outputs equally, which can dilute individual layer contributions and produce depth-wise representation redundancy in deep VLA policies.", "limitation_search_queries": [ "Attention Residuals AttnRes PreNorm dilution hidden-state growth", "residual connections transformer PreNorm gradient flow depth", "hyper-connections deep transformers residual aggregation", "Deep Residual Learning for Image Recognition residual connections" ], "bridge_to_our_method": "AttnRes supplies a drop-in way for each policy layer to selectively aggregate earlier residual outputs rather than relying on fixed unit-weight accumulation." } ] } }, "section_plan": [ { "section_title": "Abstract", "subsections": [ { "subsection_title": "Abstract Content", "content_bullets": [ "State the robotics problem: diffusion-style VLA policies are expressive but slow when closed-loop control needs many policy evaluations.", "Introduce iMF-AttnRes as a VLA imitation-learning policy combining improved Mean Flow for one-to-few-step action generation and Attention Residuals for depth-wise residual aggregation.", "Report grounded socket-insert results from experimental_log.md: best iMF-AttnRes reaches avg_reward 1513.56 and median_reward 1901.5 with about 15.122 ms inference time, compared with diffusion-style infer100 baselines at 861.53/338.291 ms and 1019.39/397.912 ms.", "Report grounded sim_transfer result: best iMF-AttnRes reaches avg_reward 526.22 and 44/100 success-like episodes compared with native diffusion policy at avg_reward 319.2 and 29/100 success-like episodes.", "Mention limitations: simulation-only, heterogeneous hardware, and sensitivity to horizon/execution settings." ], "citation_hints": [] }, { "subsection_title": "Abstract Grounding Checks", "content_bullets": [ "Ensure every numeric claim in the abstract is copied from experimental_log.md Table 1, Table 2, or Table 3.", "Do not cite papers in the abstract unless the final manuscript style requires it; preserve double-blind anonymity." ], "citation_hints": [] } ] }, { "section_title": "1. Introduction", "subsections": [ { "subsection_title": "1.1 Motivation and Problem Setting", "content_bullets": [ "Motivate fast closed-loop manipulation in RoboIMI socket-insert and sim_transfer environments using the latency and reward metrics in experimental_log.md.", "Explain that iterative diffusion-style inference can require infer100 evaluations in the baselines, producing hundreds of milliseconds per inference on socket-insert runs.", "Frame the research question: can a one-to-few-step average-flow action generator preserve or improve manipulation reward while reducing inference latency?" ], "citation_hints": [ "Chi et al. (Diffusion Policy: Visuomotor Policy Learning via Action Diffusion)", "research paper or technical report introducing 'Vision-Language-Action robot policies'" ] }, { "subsection_title": "1.2 Contributions", "content_bullets": [ "List contributions: adapting improved Mean Flow to VLA action generation, integrating AttnRes into the policy transformer, and evaluating on two RoboIMI simulated manipulation tasks.", "State that the paper provides speed-quality comparisons against diffusion-style VLA, Diffusion Policy native DiT/DDPM/ResNet, ACT, and SmolVLA baselines using the exact 100-rollout statistics from experimental_log.md.", "Emphasize honest scope: first draft simulation evidence rather than real-robot transfer." ], "citation_hints": [ "Geng et al. (Mean Flows for One-step Generative Modeling)", "Geng et al. (Improved Mean Flows: On the Challenges of Fast-Forward Generative Models)", "research paper or technical report introducing 'Attention Residuals'" ] } ] }, { "section_title": "2. Related Work", "subsections": [ { "subsection_title": "2.1 Diffusion and Flow Policies for Robot Action Generation", "content_bullets": [ "Discuss diffusion-policy-style visuomotor action generation and its reliance on iterative denoising or repeated policy evaluations.", "Connect diffusion and flow matching as continuous generative modeling tools, then distinguish iMF-AttnRes by its average-flow one-to-few-step inference objective.", "Use related-work cluster 2.1 from outline.json to bridge to the method." ], "citation_hints": [ "Chi et al. (Diffusion Policy: Visuomotor Policy Learning via Action Diffusion)", "Ho et al. (Denoising Diffusion Probabilistic Models)", "Lipman et al. (Flow Matching for Generative Modeling)", "Liu et al. (Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow)", "Peebles and Xie (Scalable Diffusion Models with Transformers)" ] }, { "subsection_title": "2.2 Vision-Language-Action Models and Action Chunking", "content_bullets": [ "Summarize transformer and VLA robot policies as the broader setting for the proposed policy.", "Discuss action chunking and compact VLA policies as complementary routes to practical control speed.", "Position iMF-AttnRes as an action-generation module that can coexist with VLA conditioning and chunked execution." ], "citation_hints": [ "Brohan et al. (RT-1: Robotics Transformer for Real-World Control at Scale)", "Brohan et al. (RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control)", "Kim et al. (OpenVLA: An Open-Source Vision-Language-Action Model)", "Zhao et al. (Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware)", "research paper or technical report introducing 'SmolVLA'" ] }, { "subsection_title": "2.3 Mean-Flow Training for One-Step Generation", "content_bullets": [ "Use the derivations in idea.md to explain the distinction between instantaneous velocity v(x_t,t) and average velocity over [r,t].", "Explain the Mean Flow relation between average velocity and instantaneous velocity and why the JVP term appears.", "Explain the improved Mean Flow correction from idea.md: using the corrected relation to retain flow-matching training stability rather than directly substituting conditional velocity inside a nonlinear JVP expression." ], "citation_hints": [ "Geng et al. (Mean Flows for One-step Generative Modeling)", "Geng et al. (Improved Mean Flows: On the Challenges of Fast-Forward Generative Models)", "Lipman et al. (Flow Matching for Generative Modeling)", "Liu et al. (Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow)" ] }, { "subsection_title": "2.4 Residual Aggregation and Attention Residuals", "content_bullets": [ "Derive the residual accumulation view in idea.md: x_t = y_0 + ... + y_t and y_{t+1}=f_{t+1}(sum y_s).", "Explain the generalization to weighted residual aggregation and the AttnRes form a_{t+1,s} proportional to exp(w_{t+1} dot RMSNorm(y_s)).", "Discuss why learned residual aggregation may help deeper policy networks avoid feature dilution and gradient imbalance." ], "citation_hints": [ "He et al. (Deep Residual Learning for Image Recognition)", "Vaswani et al. (Attention Is All You Need)", "research paper or technical report introducing 'Attention Residuals'", "research paper or technical report introducing 'PreNorm Transformers'" ] } ] }, { "section_title": "3. Method", "subsections": [ { "subsection_title": "3.1 Problem Formulation", "content_bullets": [ "Define the imitation-learning policy input as multi-view visual observations, robot state, and optional language/task conditioning from the RoboIMI environment, and define the output as an action chunk of length exec.", "Introduce the evaluation setting from experimental_log.md: socket-insert and sim_transfer 100-rollout evaluations with reward, success-like count, inference FPS, control FPS, and inference time.", "Avoid inventing dataset sizes, optimizer settings, or training details not present in idea.md or experimental_log.md." ], "citation_hints": [ "research paper or technical report introducing 'behavior cloning for robotics'", "research paper or technical report introducing 'Transformer architecture'" ] }, { "subsection_title": "3.2 Improved Mean Flow Action Generation", "content_bullets": [ "Formalize the ODE dx_t/dt = v(x_t,t) and average velocity ̄v(z_t,r,t)=1/(t-r) int_r^t v(z_tau,tau)d tau using the notation from idea.md.", "Show the Mean Flow identity ̄v = v - (t-r) d̄v/dt and the JVP expression jvp(̄v,(z,r,t),(v,0,1)).", "Explain the iMF training relation V_theta(z_t)=u_theta(z_t)+(t-r)JVP_sg(u_theta;v_theta) and that gradients update the average-flow branch u_theta, supporting one-to-few-step inference." ], "citation_hints": [ "Geng et al. (Mean Flows for One-step Generative Modeling)", "Geng et al. (Improved Mean Flows: On the Challenges of Fast-Forward Generative Models)", "Lipman et al. (Flow Matching for Generative Modeling)" ] }, { "subsection_title": "3.3 Attention Residual Policy Transformer", "content_bullets": [ "Use the residual decomposition from idea.md to show fixed residual accumulation as a uniform sum over layer outputs.", "Define AttnRes weights a_{t+1,s} with a softmax over w_{t+1} dot RMSNorm(y_s), then define the layer input as a weighted sum of previous residual outputs.", "State that the paper uses AttnRes primarily in the policy transformer in the strongest variants; the full-AttnRes vision ablation underperformed in sim_transfer." ], "citation_hints": [ "He et al. (Deep Residual Learning for Image Recognition)", "Vaswani et al. (Attention Is All You Need)", "research paper or technical report introducing 'Attention Residuals'" ] }, { "subsection_title": "3.4 Inference and Control Loop", "content_bullets": [ "Describe one-to-few inference evaluations for iMF-AttnRes, using infer1/infer2/infer3 in socket runs and infer1 in sim_transfer best run.", "Contrast this with diffusion-style infer100 baselines in experimental_log.md.", "Explain that predicted chunks are executed for exec steps, e.g., exec16 in the best socket and sim_transfer iMF-AttnRes variants." ], "citation_hints": [ "Chi et al. (Diffusion Policy: Visuomotor Policy Learning via Action Diffusion)", "Zhao et al. (Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware)" ] } ] }, { "section_title": "4. Experiments", "subsections": [ { "subsection_title": "4.1 Experimental Setup", "content_bullets": [ "Describe RoboIMI socket-insert and sim_transfer environments as simulation benchmarks using only details stated in experimental_log.md.", "Define metrics exactly as in experimental_log.md: avg_reward, median_reward, avg_max_reward, max_reward, nonzero_reward, success_like, avg_inference_fps, avg_control_fps, avg_inference_time_ms.", "State that each row is based on 100 simulation rollouts and that hardware varies across run names, so speed is practical but not fully hardware-normalized." ], "citation_hints": [ "research paper or technical report introducing 'Diffusion Policy: Visuomotor Policy Learning via Action Diffusion'", "research paper or technical report introducing 'ACT action chunking transformer'", "research paper or technical report introducing 'SmolVLA'" ] }, { "subsection_title": "4.2 Socket Peg Insertion Results", "content_bullets": [ "Create a LaTeX table from Table 1 in experimental_log.md with the seven socket policies and their reward/speed metrics.", "State that best iMF-AttnRes infer3 obtains avg_reward 1513.56 and median_reward 1901.5, exceeding diffusion-style ph16 infer100 avg_reward 861.53 and ph32 infer100 avg_reward 1019.39.", "State that iMF-AttnRes infer2 latency is 14.409 ms and infer3 latency is 15.122 ms, compared with 338.291 ms and 397.912 ms for diffusion-style infer100 baselines.", "Discuss the nonzero_reward caveat: diffusion-style ph16 has 97/100 nonzero episodes, but much lower median and average reward than iMF-AttnRes." ], "citation_hints": [] }, { "subsection_title": "4.3 Sim Transfer Results and Ablations", "content_bullets": [ "Create a LaTeX table from Table 2 in experimental_log.md summarizing native diffusion policy, diffusion-style VLA, and iMF ablations.", "Report that sim-transfer iMF-AttnRes infer1 reaches avg_reward 526.22 and 44/100 success-like episodes, compared with native diffusion policy at avg_reward 319.2 and 29/100 success-like episodes.", "Analyze ablations: multi-token ResNet18 iMF is very fast at 441.80 FPS but lower reward; full-AttnRes vision underperforms; horizon/execution choices are sensitive." ], "citation_hints": [] }, { "subsection_title": "4.4 Speed-Quality Tradeoff and Failure Modes", "content_bullets": [ "Use Table 3 from experimental_log.md for derived ratios: 1.76x socket reward over ph16 diffusion baseline, 22.97x socket inference FPS over ph16 diffusion baseline, and 1.65x sim_transfer average reward over native diffusion policy.", "Discuss failure modes from experimental_log.md: simulation-only evidence, heterogeneous hardware, threshold differences across tasks, and sensitivity to horizon/execution and vision-residual design.", "Emphasize that iMF-AttnRes improves latency but still needs controlled hardware-normalized studies and real-robot transfer validation." ], "citation_hints": [] } ] }, { "section_title": "5. Limitations", "subsections": [ { "subsection_title": "5.1 Scope and Evaluation Limits", "content_bullets": [ "Explicitly satisfy CoRL requirement by describing simulation-only evaluation and lack of real-robot experiments.", "State heterogeneous hardware makes speed comparisons imperfect and motivates future hardware-normalized benchmarking.", "State that iMF-AttnRes sensitivity to horizon/execution choices and full-AttnRes vision underperformance are current limitations." ], "citation_hints": [] }, { "subsection_title": "5.2 Future Work", "content_bullets": [ "Propose future work: real-robot transfer, more seeds, hardware-normalized latency tests, and better criteria for where to apply AttnRes in vision vs policy modules.", "Mention possible integration with efficient VLA backbones and action tokenization but avoid claiming untested results." ], "citation_hints": [] } ] }, { "section_title": "6. Conclusion", "subsections": [ { "subsection_title": "Conclusion Content", "content_bullets": [ "Summarize that improved Mean Flow plus AttnRes is a practical recipe for low-latency VLA action generation in the reported RoboIMI simulations.", "Restate the strongest socket and sim_transfer quantitative evidence without adding new numbers.", "Close with a balanced statement that this is promising simulation evidence requiring real-robot validation." ], "citation_hints": [] }, { "subsection_title": "Concluding Caveats", "content_bullets": [ "End with simulation-only and hardware-normalization caveats from experimental_log.md rather than overclaiming real-robot deployment.", "Mention future validation directions without adding unsupported experiments." ], "citation_hints": [] } ] } ] }