{"id":"24d61887-8a79-49f8-b236-2f3e23d7b250","arxiv_id":"2507.17338","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hierarchical active inference agent with a whole-body controller achieves 66.5% average success on Habitat's three long-horizon rearrangement tasks, beating the compared RL baselines (54.7%).","lead":"This paper builds a fully hierarchical active inference controller, a non learned approach, and tests it on three long horizon mobile manipulation tasks in the Habitat benchmark. It reports higher average success than two reinforcement learning baselines, but relies on privileged map and object state information and hand tuned skills.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'outperforms SOTA baselines' claim is not yet supported because the method receives privileged global-map and articulated-object-state signals the RL baselines do not; a controlled ablation is needed before the comparison can be attributed to active inference.","rationale":"I read the paper as making an empirical scaling claim, not a purely theoretical one: active inference, without offline training, beats trained RL baselines on a modern benchmark. For that claim to hold, the comparison must isolate the architecture. The manuscript's own Sec. 3.2 admits privileged information, and Sec. 2.1 replaces active-inference navigation with A* on a global map. The reader's weakest-assumption analysis identifies exactly this. I agree with the conditional verdict: the system is novel, transparent, and the HAIF derivation is plausible (closed-form gradients in Appendix A.1), but the headline quantitative claim is not yet established by the reported numbers. The lack of error bars makes the TidyHouse 1.5-point gap especially fragile. I do not see a stronger internal inconsistency in the equations; the concern is about the comparison protocol and the scope of the central claim. A controlled ablation is the cleanest way to settle it.","tokens_in":12423,"tokens_out":6639,"duration_ms":72322,"concrete_test":"Run the proposed method without the two privileged signals: replace global-map A* navigation with path planning on the online VBGS map only, and infer drawer/fridge open/closed state from raw RGBD instead of ground truth. Evaluate on the same 100 episodes per task and five seeds used in Fig. 4. If the three-task average drops by more than the current 11.8-point margin, or if the TidyHouse gap reverses, the headline outperformance cannot be attributed to active inference rather than to the privileged signals.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that hierarchical active inference outperforms state-of-the-art baselines on the Habitat Benchmark. The load-bearing condition is a fair comparison, and the paper itself flags the weak point. Section 2.1 says navigation 'opted to use A* pathfinding' using 'prior knowledge about the object location'; Section 3.2 admits 'we still rely on privileged information, such as the floor map for path planning and articulated object states.' The Multi-skill RL baseline from [10] is trained from egocentric observations and does not receive these signals. The reported advantage (66.5% vs 54.7%) may therefore reflect access to a global floor plan and ground-truth drawer/fridge states rather than the active inference architecture. The per-task numbers support this concern: TidyHouse is only 72.5% vs 71% (1.5 points, within plausible noise given no error bars are reported), while the largest gap is SetTable (50% vs 29%), exactly the task requiring knowledge of articulated object states. The manuscript also does not state explicitly whether the MM baseline was rerun on the same episodes or taken from [10]; if the latter, episode-level comparability is an additional confound. This is a limitation of the evidence, not an internal inconsistency, and the authors disclose the privileged signals and promise to remove them.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical active inference architecture for long-horizon mobile manipulation and evaluates it on three Habitat Benchmark tasks: TidyHouse, PrepareGroceries, and SetTable. The architecture combines a discrete high-level POMDP planner, a navigation module based on A* pathfinding, a Variational Bayes Gaussian Splatting (VBGS) world map, and a continuous whole-body hierarchical active inference controller. The authors report average success rates of 72.5%, 77%, and 50% on the three tasks, compared with 71%, 64%, and 29% for the Multi-skill RL mobile manipulation baseline from Gu et al. (2022), and an aggregate success rate of 66.5% versus 54.7%. They conclude that this is the first demonstration that active inference can scale to modern robotics benchmarks. The paper explicitly discloses reliance on privileged information, namely a global floor map for path planning and ground-truth articulated object states, and states that these assumptions will be removed in future work.","tokens_in":12785,"tokens_out":5985,"duration_ms":64750,"significance":"If the empirical comparison were clean, this result would be significant: a hand-tuned, non-trained hierarchical active inference controller outperforming RL-based mobile manipulation policies on a standard benchmark would challenge the prevailing assumption that long-horizon rearrangement tasks require extensive offline training. The paper also makes a useful technical contribution by extending hierarchical active inference to whole-body differential-drive control, particularly the coupling of base and arm prediction errors in Eq. (12), and by demonstrating online probabilistic mapping with VBGS in a benchmark setting. The authors are transparent about the privileged information and the manual tuning involved, which is a strength. However, the central claim that the method 'outperforms state-of-the-art baselines' is not yet supported because the baselines do not receive the same oracle information, the baseline evaluation protocol is incompletely specified, and the sensitivity to manually chosen parameters is not assessed.","major_comments":[{"comment":"The central comparison is confounded by privileged information. In Sec. 2.1 the navigation component uses A* pathfinding with 'prior knowledge about the object location,' and Sec. 3.2 concedes that the method 'still rely[ies] on privileged information, such as the floor map for path planning and articulated object states.' The Multi-skill RL baseline from [10] is trained from egocentric observations and does not receive these oracle signals. Consequently, the reported aggregate advantage (66.5% vs. 54.7%) and the per-task differences in Fig. 4 cannot be attributed to the active inference architecture itself. Please add a controlled comparison in which the RL baselines receive the same global map and articulated-object-state information, or remove these privileged inputs from the proposed method, and report both configurations.","section":"Sec. 3.2 and Sec. 2.1"},{"comment":"The evaluation protocol is not sufficiently specified for a fair comparison. The manuscript does not state whether the MM and Monolithic RL results were rerun by the authors on the same 100 episodes or taken directly from [10]; if the latter, episode-level comparability is not guaranteed. In addition, no error bars, confidence intervals, or per-seed breakdowns are reported, and the text says results are averaged 'over five seeds' while Fig. 4 says 'averaged over 100 episodes.' Please reconcile these statements, report variance across seeds, and describe exactly how each baseline was evaluated.","section":"Sec. 3.2, Fig. 4"},{"comment":"The claim that the method requires no offline training is qualified by a large set of manually chosen parameters. App. A.2 lists repulsion and threshold gains (kr,theta, kr,obst, gamma_theta, gamma_obst) and states that parameters are 'manually chosen to achieve sufficient performance in the test cases'; Sec. 2.4 and Sec. 3.2 refer to hand-tuned skill sequences and goals. Since the reported success rates may depend strongly on these choices, please report the tuning budget, the sensitivity of the results to parameter variations, or both, so that the generalization claim can be assessed.","section":"App. A.2 and Sec. 2.4"},{"comment":"The statement that the method 'outperforms state-of-the-art baselines' is based on a comparison with only two methods from [10]. The Habitat Benchmark paper [28] itself proposes a hierarchical baseline, and other published results may exist on these tasks. Please either include a broader set of baseline methods or justify explicitly why the two methods from [10] constitute the current state of the art on this benchmark.","section":"Sec. 3.1, Baselines"}],"minor_comments":[{"comment":"Equation (17) prints the same expression for all four quaternion components; based on the orientation expressions that follow, the components should be [cos(theta/2)cos(alpha/2), cos(theta/2)sin(alpha/2), sin(theta/2)sin(alpha/2), sin(theta/2)cos(alpha/2)]. Please correct this typo, as it affects the reproducibility of the kinematic model.","section":"App. A.1, Eq. (17)"},{"comment":"The heading 'Average Sucess Rate' contains a typo; it should read 'Average Success Rate.'","section":"Fig. 4"},{"comment":"The first sentence of App. A.1 reads 'To generative model in eq. (4)'; it should read 'The generative model in Eq. (4).'","section":"App. A.1"},{"comment":"The Navigation Model is implemented with A* pathfinding, which is not an active inference component. The abstract and introduction should describe the system as a hybrid active inference architecture that includes classical planning, to avoid overclaiming that the entire system is active inference.","section":"Sec. 2.1"},{"comment":"The transition from results to the statement 'demonstrating for the first time that active inference can scale to the complexity of modern robotics benchmarks' is too strong given the privileged information and the narrow baseline set; please qualify the conclusion accordingly.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a plausibly important empirical claim for the active inference community, but the headline result depends on a comparison that is currently confounded by oracle information given only to the proposed method. The authors are transparent about this limitation, and the central architecture is interesting enough that the claim could be supported by a controlled ablation in a revision. I would not reject the paper, but the empirical section needs substantial strengthening before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Rough read: the paper gives active inference a credible shot at a modern benchmark, but the headline 'outperforms baselines' is weaker than it looks because the agent gets the floor plan and ground-truth drawer/fridge states while the RL baselines do not. That said, there is real substance here.\n\nWhat is actually new: the differential-drive generative model (eq. 9) and the whole-body coupling update (eq. 12) are genuine extensions of HAIF to a mobile base plus arm. The high-level discrete planner over skills is a sensible way to get long-horizon behavior. Running this on Habitat and getting non-trivial success rates without any offline training is a first for active inference, and the paper deserves credit for reporting the numbers honestly, including the privileged-info caveat. The math in the appendix is coherent; the closed-form gradients look right.\n\nThe soft spot is the comparison. The 66.5% vs 54.7% average is the centerpiece, but it is confounded. The method uses A* on a global map and knows when drawers and fridges are open. The RL baselines from Gu et al. operate from egocentric observations and do not get those signals. The per-task pattern supports the concern: the smallest gap is TidyHouse (1.5 points), the largest is SetTable (21 points), which is exactly the task that depends on knowing articulated-object states. No error bars are reported, and it is not stated whether the MM baseline was rerun on the same episodes or taken from the paper. That last point matters for episode-level comparability. So the claim 'outperforms state-of-the-art baselines' is not yet established; the evidence supports 'competitive, under privileged information.'\n\nThe paper's own discussion acknowledges these limitations, which is why I would not call this a fatal flaw. It is a proof-of-concept with an overreaching abstract. For the active inference community, it shows the framework can handle a complex whole-body task; for robot learning, it is a reminder that benchmark comparisons need controlled inputs.\n\nThis deserves a serious referee. The right revision would add a no-privileged-info ablation, include a couple more baselines or at least the Habitat leaderboard numbers, add standard deviations, and release code or a detailed parameter appendix.","headline":"Real extension of HAIF to whole-body mobile manipulation, but the benchmark claim is confounded by privileged information.","tokens_in":13294,"tokens_out":2168,"would_cite":true,"duration_ms":22633,"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":"A fully hierarchical active inference agent outperforms trained RL baselines on long-horizon mobile manipulation tasks in the Habitat benchmark, reaching 66.5% average success versus 54.7%.","keywords":["active inference","mobile manipulation","long-horizon task","hierarchical control","whole-body control","Habitat benchmark","free energy principle","variational Bayes Gaussian splatting"],"falsifier":"Give the Multi-skill RL baseline the same global map and oracle knowledge of articulated object states, rerun the 100-episode evaluation on the three tasks, and check whether its combined success rises to match or beat 66.5%; or remove those privileges from the active inference agent and check whether its success falls to or below 54.7%. Either outcome would settle whether active inference itself, rather than the extra information, is responsible for the reported outperformance.","tokens_in":12280,"feed_emoji":"🤖","tokens_out":9103,"duration_ms":83685,"temperature":0.7,"pith_summary":"The paper sets out to show that active inference—a framework in which perception, planning, and control all minimize the same variational free energy—can handle long-horizon robot tasks that currently demand trained reinforcement learning policies. It introduces a fully hierarchical active inference agent in which a high-level discrete model sequences skills such as Pick, Place, and Move, and a continuous whole-body controller generates joint- and base-level commands. Evaluated on three mobile manipulation tasks from the Habitat benchmark (TidyHouse, PrepareGroceries, and SetTable), the agent reaches a combined success rate of 66.5% against 54.7% for the best RL baseline, without any offline training. The authors interpret this as the first demonstration that active inference scales to the complexity of modern robotics benchmarks.","feed_headline":"Active inference beats trained RL on long-horizon robot tasks","feed_subtitle":"Fully hierarchical controller hits 66.5% on Habitat benchmark, beating 54.7% RL baseline without offline training.","key_machinery":"The central object is a hierarchy of generative models, each minimizing variational free energy, with actions of one level becoming the preferences of the level below. At the top, a partially observed Markov decision process sequences discrete skills (Pick, Place, Move, PickFromDrawer, PickFromFridge) by tracking the robot's relation to pick and place locations and the object's inventory state; a retry model switches approach directions on pick/place failure. The Navigation model uses A* pathfinding over a probabilistic world map built online with Variational Bayes Gaussian Splatting (VBGS), which represents free and occupied space as a 6D Gaussian mixture over 3D points with color. At the bottom, a whole-body hierarchical active inference (HAIF) controller solves for joint- and base-level controls: the arm uses a quaternion-based kinematic generative model, the mobile base uses a differential-drive generative model, and the two are linked by propagating base and arm extrinsic prediction errors into each other's belief updates, letting the base extend the arm's reach by moving to minimize the arm's errors. Goals act as linear attractors and obstacles as repulsive forces in the dynamics function, and the control signal minimizes the proprioceptive component of free energy.","core_discovery":"On the paper's own terms, the central claim is that a fully hierarchical active inference architecture outperforms state-of-the-art baselines on all three long-horizon tasks of the Habitat benchmark. The agent combines a high-level active inference model that selects among discrete whole-body skills with a continuous whole-body controller based on hierarchical active inference, in which the mobile base and the arm minimize each other's prediction errors. Over five seeds and 100 episodes per task, the method achieves 72.5% success on TidyHouse, 77% on PrepareGroceries, and 50% on SetTable, for a combined 66.5%, versus 71%, 64%, and 29% (54.7% combined) for the multi-skill RL baseline. The method requires no offline training; skills are hand-tuned over a handful of episodes, and the agent can recover from pick and place failures online by switching approach directions. The paper acknowledges that it currently relies on privileged information—a global floor map for path planning and ground-truth knowledge of drawer and fridge states—which the baselines do not receive, and states that these assumptions will be removed in future work.","pith_inferences":["If the performance gap persists under equal information, the practical consequence is that composing cheap, hand-tuned model-based skills with online replanning is a data-efficient alternative to large RL pipelines for rearrangement, not just on this benchmark but on comparable long-horizon tasks.","The paper's own admission that it uses a global map and oracle object states implies perception is currently the bottleneck; a natural extension is measuring how success degrades as each privilege is removed, or how it rises when a baseline is given the same privileges.","The base-arm coupling mechanism is a generic control principle that could transfer to other redundant mobile manipulators; a direct test would be to port the whole-body controller to a different platform and compare reach and success on the same skills.","The authors hint at adding an intermediate hierarchical level to smooth transitions between subgoals; if that layer were learned, the architecture would test whether a learned skill level can combine with the hand-tuned core to push success further."],"forward_implications":["Long-horizon mobile manipulation no longer necessarily requires offline RL training: the skills are hand-tuned and the high-level plan is recomputed online, avoiding reward shaping and millions of training steps.","Failures can be absorbed online: the retry model switches approach parameters on pick/place failures and the high-level model can re-sequence remaining subtasks, unlike the open-loop chained-RL baseline.","The whole-body coupling lets the base absorb the arm's prediction errors, extending the arm's reachable workspace and relaxing the need to dock the base precisely before manipulation.","The 11.8-point combined success advantage indicates that free-energy-minimizing model-based control is competitive with trained policies on established robotics benchmarks, rather than limited to simple 2D settings."],"supporting_citations":[{"why":"Defines the Habitat Benchmark's three long-horizon tasks, the evaluation protocol, and the hierarchical framework the paper builds on.","marker":"[28]"},{"why":"Supplies the Multi-skill RL and Monolithic RL baselines that the method compares against, as well as the mobile-manipulation skill structure that is adapted.","marker":"[10]"},{"why":"Provides the hierarchical active inference belief-update equations and the deep kinematic generative model at the core of the continuous controller.","marker":"[22]"},{"why":"Extends hierarchical active inference to robot arm control, which the paper extends to whole-body control of a mobile manipulator.","marker":"[20]"},{"why":"Introduces Variational Bayes Gaussian Splatting, the online probabilistic map representation that supplies obstacle and goal information to the hierarchy.","marker":"[15]"},{"why":"A* pathfinding generates waypoints for the Navigation model, turning goals into low-level extrinsic targets.","marker":"[11]"},{"why":"Sets out the free energy principle and active inference formalism that justifies both the discrete high-level model and the continuous controller.","marker":"[18]"}],"fun_headline_variants":["Active inference tops RL on long-horizon manipulation","Hierarchical active inference beats multi-skill RL on Habitat","No offline training: active inference wins long-horizon tasks","Active inference scales to realistic long-horizon benchmarks","Better than RL on 3 long-horizon tasks with active inference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the measured advantage comes from the active inference architecture itself and not from privileged information: the agent is given a global floor map for path planning and ground-truth drawer/fridge states that the RL baselines do not receive; if a baseline were given the same information, or the agent had to operate from raw RGBD, the reported gap could shrink, vanish, or reverse.","fun_headline_variants_meta":{"raw":{"variants":["Active inference tops RL on long-horizon manipulation","Hierarchical active inference beats multi-skill RL on Habitat","No offline training: active inference wins long-horizon tasks","Active inference scales to realistic long-horizon benchmarks","Better than RL on 3 long-horizon tasks with active inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000345,"raw_usage":{"total_tokens":1866,"prompt_tokens":893,"completion_tokens":973,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":890}},"tokens_in":509,"tokens_out":973,"duration_ms":10162,"temperature":1.0,"reasoning_tokens":890,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:50:00.533151+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give the Multi-skill RL baseline the same global map and oracle knowledge of articulated object states, rerun the 100-episode evaluation on the three tasks, and check whether its combined success rises to match or beat 66.5%; or remove those privileges from the active inference agent and check whether its success falls to or below 54.7%. Either outcome would settle whether active inference itself, rather than the extra information, is responsible for the reported outperformance.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Habitat Benchmark's three long-horizon tasks, the evaluation protocol, and the hierarchical framework the paper builds on."},{"cited_title":"In: The First Workshop on NeuroAI@ NeurIPS2024 (2024)","cited_arxiv_id":null,"evidence_quote":"Extends hierarchical active inference to robot arm control, which the paper extends to whole-body control of a mobile manipulator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces Variational Bayes Gaussian Splatting, the online probabilistic map representation that supplies obstacle and goal information to the hierarchy."},{"cited_title":"MIT Press (2022)","cited_arxiv_id":null,"evidence_quote":"Sets out the free energy principle and active inference formalism that justifies both the discrete high-level model and the continuous controller."}],"review_version":1}