{"id":"bc3fad99-aa8a-4298-bd53-73dc72071078","arxiv_id":"2411.13020","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AsymDex trains two multi-fingered robot hands for bimanual tasks by assigning asymmetric roles and using relative coordinates, beating baselines in success and sample efficiency.","lead":"This paper presents AsymDex, a reinforcement learning system that teaches two robotic hands to work together on tasks like opening a bottle, by giving each hand a different job (one holds and tilts, the other twists). It reports that this asymmetric, coordinate-centric approach learns faster and succeeds more often than previous methods, in simulations and on real robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sec. 4.2 removes the facilitating hand's absolute pose, but in a gravity field that pose determines the gravity direction in the relative frame, so states with identical AsymDex observations can require different actions.","rationale":"I read the paper as claiming that AsymDex efficiently learns a broad class of asymmetric bimanual dexterous tasks without demonstrations, with the two key design choices being asymmetric hand roles and relative observation/action spaces. The strongest evidence is the consistent simulation wins over the ablations and the real-world zero-shot deployments. My concern targets the relative-coordinate design itself: the derivation in Sec. 4.2 removes ξ^b_f on the grounds that the facilitating hand does not move relative to its held object, but the absolute orientation of that object is still needed to know which way gravity acts in the relative frame. This is an internal representational issue, not a disagreement with community consensus, and it is more specific than the reader's rigidity concern. Even under the paper's own rigidity assumption, the reduced observation space is not information-complete for gravity-dependent tasks such as placing, stacking, pouring, and lid-twisting. The proposed test directly varies the facilitating-held object's initial orientation; if AsymDex still succeeds, the concern is refuted and the relative-space design is robust. If it fails, the central claim needs to be narrowed or the observation space must include the absolute orientation of the held object. I keep the reader's CONDITIONAL verdict because the issue is addressable with additional experiments, and the existing results on the tested distributions may remain valid, but the condition should now explicitly include observational sufficiency under varied facilitating-hand orientations.","tokens_in":16960,"tokens_out":10861,"duration_ms":118533,"concrete_test":"Run the Bottle cap and Pour interaction-phase experiments with the initial orientation of the facilitating-held object sampled uniformly over yaw in [0, 2π) (and a moderately larger pitch range), while keeping AsymDex's observation/action spaces and hyperparameters unchanged. Compare success rate and sample efficiency against (a) AsymDex with the original narrow orientation ranges and (b) an oracle variant whose observation additionally includes the held object's absolute orientation or the gravity direction in P_f. If the wide-orientation variant collapses while the oracle succeeds, the removed ξ^b_f in Sec. 4.2 is information-bearing and the 'without losing any information' claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the claim in Sec. 4.2 that, because there is no relative motion between the facilitating hand and its held object, the facilitating hand base pose and held-object pose can be removed from the observation space 'without losing any information.' This is only valid if the dynamics and rewards are invariant to the absolute pose of the frame P_f. In all four tasks, gravity breaks this symmetry: the direction of gravity in P_f is exactly the information carried by the removed ξ^b_f. With z_AsymDex = [ξ^h_d, ξ^b_r, ξ^{obj}_r], two states that differ only by rotating the facilitating hand and its held object together (e.g., 90 degrees about the world vertical) produce identical observations but different optimal actions: releasing a block 'down' relative to the cup in one state versus 'sideways' in the other. Since the policy is an MLP with no recurrence and only the current observation plus previous action, it cannot infer the missing gravity direction from history. The empirical success may therefore depend on the very narrow initial-orientation ranges in Appendix C (e.g., facilitating hand rotation ±0.5 rad for Bottle cap, ±0.1 rad for the real Pour task), rather than on the sufficiency of the relative representation. This is distinct from the rigidity assumption flagged by the reader: even with a perfectly rigid grasp, the removed ξ^b_f is information-bearing in a gravitational field. If this concern lands, the breadth of the 'large class' claim and the relative-space ablation conclusions are weakened, though the comparisons on the tested narrow distributions could still stand.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents AsymDex, a reinforcement-learning framework for bimanual dexterous manipulation with multi-fingered hands. The method assigns asymmetric roles to the two hands—a facilitating hand that holds and reorients one object and a dominant hand that performs the fine manipulation—and defines the policy's observation and action spaces in the coordinate frame of the object held by the facilitating hand, together with a relative-pose tracking controller that converts relative target actions into absolute base motions. The paper also proposes a two-phase decomposition that separates object acquisition (grasping) from the coordinated interaction phase. AsymDex is evaluated in four simulated tasks (Block in cup, Stack, Bottle cap, Switch) and three real-world tasks (Block in cup, Pour, Twist Lid) against baselines that ablate the asymmetry and relative-space choices, and the authors report higher success rates and sample efficiency in simulation, plus zero-shot sim-to-real deployment on hardware.","tokens_in":17269,"tokens_out":15319,"duration_ms":153172,"significance":"If the results hold, AsymDex is a useful and simple framework for a practically important class of asymmetric bimanual manipulation tasks, and the simulation study is systematic: four tasks, five seeds, and three ablations that challenge each design choice. The paper also demonstrates zero-shot transfer for the proposed method on three real-world tasks. However, the theoretical justification for the observation-space reduction is incomplete, the key controller hyperparameter α is not reported, and some of the reported 'consistent' wins are within large error bars. The rigid-grasp scope is acknowledged in Sec. 7, but the gravity dependence of the relative representation is not. With these issues addressed, the paper would be a solid contribution.","major_comments":[{"comment":"The claim that ξ^b_f(t) and ξ^obj_f(t) can be removed from the observation 'without losing any information' is not valid in the presence of gravity. The observation z_AsymDex is invariant under a global rotation of the entire scene, but the dynamics and the rewards are not: the Pour reward R_cup_orient = z_cupf · z_world (Appendix D) explicitly depends on the absolute orientation of the facilitating hand's frame, and the Block-in-cup success criterion requires the cup to be oriented so that the released block falls into it. Two states that differ by a rotation of the whole scene about a horizontal axis produce identical observations but require different relative actions, because the direction of gravity in P_f differs. The controller in Eq. (1) provides the current frame orientation Rof_world, but it does not provide the direction of gravity in that frame, and the memoryless MLP policy with only the current observation and previous action cannot recover it. This is a genuine loss of information, not merely a restriction to tasks with rigid grasps; Sec. 7 lists other limitations but does not mention this gravity dependence. Please either add a gravity-direction vector expressed in P_f to the observation, or substantially qualify the losslessness claim and the 'large class' statement.","section":"Sec. 4.2"},{"comment":"The hyperparameter α is central to the bimanual controller, but its value is never reported and no ablation or sensitivity analysis is provided. The controller's behavior changes qualitatively with α: for α = 1 the facilitating hand base is stationary, while for α = 0.5 both hands share the relative motion. Without knowing α for each task, including whether it was tuned per task, the results in Tables 1–3 are not fully reproducible, and the comparison against the baselines may be affected by this choice. Please report α per task and include a sensitivity study.","section":"Sec. 4.3, Eq. (1)"},{"comment":"The statement that AsymDex 'consistently outperforms' all baselines is only clearly supported for Block in cup and Switch. For Stack, Rel-w/o-Asym reaches 0.6560 ± 0.3213 versus 0.8392 ± 0.0596 for AsymDex, and for Bottle cap, Rel-w/o-Asym reaches 0.4730 ± 0.2011 versus 0.6295 ± 0.1422. With five seeds, these differences are within overlapping error bars and should not be described as consistent wins without a statistical test or per-seed reporting. Please add confidence intervals or p-values, or rephrase the conclusion to separate the tasks with large margins from those with small ones.","section":"Sec. 5.1.1, Table 1"}],"minor_comments":[{"comment":"Lines 7 and 16 use the undefined symbols ξ^b_a(t) and ξ^h_r(t); presumably these should be ξ^b_d(t) and ξ^h_d(t).","section":"Algorithm 1"},{"comment":"The appendix title says 'BiDexHands' while the main text uses 'BiDexHand'; please unify the terminology.","section":"Appendix C"},{"comment":"Because only AsymDex policies were deployed on hardware, the real-world section should explicitly state that it is a transfer demonstration, not a comparison, and the main text should say so as well.","section":"Sec. 5.2"},{"comment":"The grasping policy for the 2-stage-sym baseline uses the relative observation ξ^b_f(t) ⊖ P_f, which appears to use a relative coordinate; please clarify whether this is shared infrastructure and not part of the baseline's 'no relative spaces' condition.","section":"Appendix B"},{"comment":"The pose difference dist(·) is not defined; please specify how translation and rotation are combined (e.g., quaternion log) and how the result is weighted by α.","section":"Eq. (1)"},{"comment":"The method uses pre-grasp pose annotations for both AsymDex and the baselines; the claim of learning 'without relying on demonstrations' should be qualified to clarify that task-specific pre-grasp knowledge is used.","section":"Sec. 4.4 and Appendix B"}],"recommendation":"major_revision","confidential_remarks":"This is a promising paper whose empirical core is likely publishable, but the theoretical claim in Sec. 4.2 needs correction and the experiments need more statistical transparency. I would not reject on the gravity issue alone, since the fix (adding a gravity vector or narrowing the claim) is within scope. I would also ask the authors to confirm during revision whether α was tuned per task and to report the chosen values."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhaodong and colleagues, here's my read after going through AsymDex carefully. The empirical skeleton is solid: four simulated tasks, three real-world deployments, ablations for both design choices, five seeds. The consistent win margins (e.g., 0.77 vs 0.11 for Block in cup) are hard to explain away, and the two-phase decomposition with a learned grasping policy is a sensible extension. The components are not individually new—relative spaces and role asymmetry appear in prior gripper work—but this is the first time they're combined for multi-fingered hands, and the paper makes a clear case that both matter.\n\nThe soft spot that keeps me honest is the observation space's sufficiency. Section 4.2 says removing the facilitating hand's absolute pose and the held object's pose 'doesn't lose any information.' That's not right in a gravitational field. The gravity vector in the facilitating-object frame is exactly the information carried by the removed absolute pose. Two states that differ only by rotating the whole system about the world vertical produce identical AsymDex observations but require different actions—releasing a block 'down' vs 'sideways.' Since the policy is an MLP with no recurrence, it can't infer the missing direction from history. The reported successes likely depend on the narrow initial-orientation ranges in Appendix C (e.g., ±0.5 rad for Bottle cap, ±0.1 rad for real Pour). That means the 'large class' claim is overstated: the method works when the facilitating hand stays near a reference orientation, not across arbitrary orientations. This is distinct from the rigidity assumption the authors do flag; even with a perfectly rigid grasp, the missing absolute pose is information-bearing.\n\nOther concerns are smaller. Alpha and per-task reward weights aren't reported, so the controller's behavior is not fully reproducible. Real-world evaluations deploy only AsymDex, so the baselines' sim-to-real gap is untested. No code or data release. These are addressable.\n\nOn balance: the central idea is sound and the empirical comparison on the tested distributions is convincing, but the generality claim needs to be reined in and the gravity issue needs an explicit discussion or a fix (e.g., include gravity direction in the observation, or restrict the claim). I'd send this to peer review with a request for major revision. It's a serious piece of work that would benefit from a careful referee.","headline":"AsymDex makes a solid empirical case for asymmetric roles and relative coordinates, but its claim that the reduced observation space loses no information is false in a gravity field.","tokens_in":17799,"tokens_out":5293,"would_cite":true,"duration_ms":50574,"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":"The paper argues that assigning complementary roles to two hands and controlling relative motion between them makes reinforcement learning efficient for bimanual dexterous manipulation, with zero-shot transfer to real hardware.","keywords":["asymmetric bimanual manipulation","dexterous manipulation","multi-fingered hands","reinforcement learning","relative coordinates","sim-to-real transfer","two-phase learning","role-differentiated bimanual manipulation"],"falsifier":"A direct falsifier: run AsymDex on a task where the holding hand must reorient the object inside its grasp, such as turning a bolt while the other hand threads a nut, and measure success against a version that retains the holding hand's finger joints; if the retained-joint version succeeds while AsymDex fails, the no-relative-motion premise is refuted.","tokens_in":16761,"feed_emoji":"🖐️","tokens_out":6321,"duration_ms":65494,"temperature":0.7,"pith_summary":"This paper claims that two simple structural choices make reinforcement learning practical for bimanual tasks with multi-fingered hands, without needing demonstrations. First, instead of treating both hands symmetrically, assign one hand a fixed holding and reorienting role and the other the role of fine manipulation. Second, express the manipulator's pose relative to the object held by the holder, and split relative-pose commands between the two hands using a tracking controller. These choices shrink the observation and action spaces, and the paper reports strong gains in success rate and sample efficiency over baselines that remove either choice. If true, the framework extends demonstration-free RL to a broad class of everyday bimanual skills.","feed_headline":"One hand holds, one hand works: RL learns bimanual skills","feed_subtitle":"AsymDex halves state and action dimensions and beats symmetric baselines across four simulated and three real tasks.","key_machinery":"The load-bearing mechanism is the combination of a role-asymmetric observation and action space with a relative-pose tracking controller. The policy observes the dominant hand's joint state plus the relative poses of the dominant hand base and the manipulated object with respect to the object held by the facilitating hand, and it outputs a target relative pose plus dominant-hand joint targets. The controller distributes the commanded relative pose between the two hand bases using a mixing coefficient, making synchronization a property of the coordinate frame rather than a behavior the policy must discover. This removes the facilitating hand's fingers from the learning problem and dramatically lowers dimensionality while coupling the hands through a shared relative target.","core_discovery":"AsymDex's central claim is that asymmetric bimanual dexterity—one hand holds and reorients an object while the other performs the fine manipulation—can be learned by reinforcement learning purely from reward, provided the policy operates in a reduced asymmetric space expressed in a frame attached to the held object. The paper defines a facilitating hand whose finger joints are omitted because there is no relative motion between it and the object it grasps, and a dominant hand whose full joint state and target pose are defined relative to the facilitating hand's object. A bimanual controller then converts the commanded relative pose into absolute base targets for both hands. In four simulated tasks (Block in cup, Stack, Bottle cap, Switch), AsymDex reaches success rates of 0.77, 0.84, 0.63, and 0.67, consistently beating baselines that drop either the asymmetry or the relative spaces. The same policies transfer zero-shot to physical hardware with 16/20, 17/20, and 18/20 successes on Block in cup, Pour, and Twist Lid, and the method can be combined with a learned grasping policy in a two-phase scheme that covers tabletop acquisition.","pith_inferences":["The paper leaves implicit that the same relative-frame idea could extend beyond two hands to multi-agent coordination, since the key step is expressing one agent's goal relative to an object another agent stabilizes.","A natural ablation for future work is keeping the dimensions matched while removing only the relative coordinates, which would isolate whether the gains come from the coordinate frame or just from reduced dimensionality.","The role assignment is prescribed rather than learned; a testable extension would let the policy discover which hand should serve as the facilitator for a given task.","The rigidity assumption suggests a hard boundary: tasks requiring in-hand reorientation by the holding hand would need a different representation, so the claimed class is best understood as tasks with no relative motion between the facilitating hand and its object."],"forward_implications":["Demonstration-free RL becomes viable for a class of asymmetric bimanual tasks with multi-fingered hands, not just for a single handcrafted skill.","Observation and action dimensions drop substantially (from 176 to 88 and from 52 to 26 on the simulated tasks), which directly improves sample efficiency.","A two-phase decomposition lets the method combine learned grasping with the interaction policy, covering full tabletop initialization rather than only pre-grasped objects.","Sim-trained AsymDex policies transfer zero-shot to real hardware with high success rates on three physical bimanual tasks."],"supporting_citations":[{"why":"Supplies the asymmetric-division-of-labor account of human bimanual action that motivates assigning different roles to the two hands.","marker":"[21]"},{"why":"Provides the BiDexHand benchmark and task designs that the simulated evaluation builds on and adapts.","marker":"[14]"},{"why":"Prior lid-twisting method whose reward design and fixed-base assumption AsymDex compares against in the real-world twist-lid task.","marker":"[19]"},{"why":"Proximal-policy-optimization algorithm used to train the AsymDex policy and all baselines.","marker":"[49]"},{"why":"Pre-grasp-annotated grasping policy integrated into the acquisition phase of the two-phase decomposition.","marker":"[50]"},{"why":"Earlier use of a projected relative space for bimanual manipulation that motivates the relative coordinate design.","marker":"[26]"}],"fun_headline_variants":["AsymDex: RL bimanual dexterity without demos","Asymmetry makes bimanual RL sample-efficient","One hand holds, the other works: RL bimanual skill","Split hand roles to halve RL state space","Bimanual RL with a dominant and helper hand"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes the facilitating hand holds its object rigidly with no relative motion between that hand and the object, so the holding hand's finger joints can be dropped from the observation and action spaces; if a task requires in-hand reorientation by the holding hand, the reduced spaces cannot represent it.","fun_headline_variants_meta":{"raw":{"variants":["AsymDex: RL bimanual dexterity without demos","Asymmetry makes bimanual RL sample-efficient","One hand holds, the other works: RL bimanual skill","Split hand roles to halve RL state space","Bimanual RL with a dominant and helper hand"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1503,"prompt_tokens":1089,"completion_tokens":414,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":705,"completion_tokens_details":{"reasoning_tokens":331}},"tokens_in":705,"tokens_out":414,"duration_ms":4960,"temperature":1.0,"reasoning_tokens":331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:55:00.114167+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct falsifier: run AsymDex on a task where the holding hand must reorient the object inside its grasp, such as turning a bolt while the other hand threads a nut, and measure success against a version that retains the holding hand's finger joints; if the retained-joint version succeeds while AsymDex fails, the no-relative-motion premise is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the asymmetric-division-of-labor account of human bimanual action that motivates assigning different roles to the two hands."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the BiDexHand benchmark and task designs that the simulated evaluation builds on and adapts."},{"cited_title":"Dasari, A","cited_arxiv_id":null,"evidence_quote":"Pre-grasp-annotated grasping policy integrated into the acquisition phase of the two-phase decomposition."}],"review_version":1}