{"id":"2b5d9238-a8d1-4cb7-8666-9be9ced3c3a2","arxiv_id":"2505.06584","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"JAEGER decouples upper- and lower-body control of a humanoid robot into two jointly trained policies and reports lower pose- and velocity-tracking errors than prior whole-body controllers in simulation.","lead":"JAEGER is a humanoid control system that uses two separate controllers, one for the lower body and one for the upper body, instead of a single network. The authors report improved tracking accuracy over existing whole-body controllers in simulation and demonstrate the system on a real Unitree H1-2 robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparison is confounded by reward and curriculum differences: the claimed SOTA superiority in Section 4.2 may reflect reward engineering rather than the dual-level architecture, and the internal ablations do not resolve this.","rationale":"The reader identified the reward/curriculum confound as the weakest assumption, and I agree that this is the single most load-bearing concern. The paper's central claim is explicitly comparative: JAEGER outperforms state-of-the-art methods. The evidence for that claim rests on comparisons against baselines trained with different rewards and curricula. Because the reward design is extensive and acknowledged as requiring significant engineering, the external comparison cannot by itself attribute the performance gap to the dual-level architecture. The internal ablations are well-designed to isolate the architecture with a fixed reward, and they support the more modest claim that the dual-level controller plus supervised initialization improves tracking relative to single-controller and no-initialization variants under JAEGER's reward. However, they do not address the SOTA comparison. The concrete test I propose would settle whether the reward confound explains the external gap. Until that test is run, the appropriate verdict is conditional acceptance: the internal claims are plausible and supported, but the headline SOTA-superiority claim requires an additional controlled comparison. I have not identified a more fundamental flaw; the method is technically coherent, the ablations are internally consistent, and the limitations are disclosed. The only adjustment beyond the reader's verdict is a slightly sharper emphasis that the real-world superiority claim is currently unquantified, but this does not change the conditional recommendation.","tokens_in":12794,"tokens_out":2700,"duration_ms":30875,"concrete_test":"Retrain one external baseline, e.g., HumanPlus, with JAEGER's exact reward function (Tables 1, 11, 12) and its supervised-initialization curriculum (Section 3.3), while keeping HumanPlus's original network architecture and observation design. Use the same AMASS training/evaluation split, retargeting preprocessing, and five random seeds, then recompute the Table 2 and Table 3 metrics. If HumanPlus's errors close to within noise of JAEGER's, the reported superiority is largely reward-driven; if the gap persists, the architectural claim is supported. A complementary check is to train JAEGER Single with HumanPlus's reward and compare to JAEGER with the same reward, isolating architecture from reward.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing claim in Section 4.2 is that JAEGER 'significantly outperforms the baseline' in velocity and joint-angle tracking against HumanPlus, ExBody, and OmniH2O (Tables 2 and 3). This comparison does not control for reward function or training curriculum. JAEGER uses a heavily tuned reward (Tables 1, 11, 12) with component weights ranging from 1 to 10,000, plus a multi-stage supervised-initialization curriculum (Section 3.3). The baselines were trained with their own reward designs and curricula; the paper does not retrain them with JAEGER's reward. Consequently, the reported gap (e.g., E_upper-j 0.1093 vs 0.2187 on H1 in Table 2) may be attributable to reward engineering rather than to the dual-level controller. The internal ablations (JAEGER Single and JAEGER w.o. Init) share JAEGER's reward and therefore support the architectural and curriculum claims, but they do not isolate the architecture from the reward in the external comparison. The paper's own Limitations section (Section 6) concedes that 'the current framework requires extensive reward engineering,' which heightens this concern. A related gap is that the abstract claims superiority in real environments, but Section A.1 and the demos provide no quantitative hardware metrics, so the sim-to-real superiority claim is currently unquantified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents JAEGER, a whole-body controller for adult-size humanoids that decouples upper- and lower-body control into two policies (a lightweight MLP for the upper body and a Gated Transformer-XL for the lower body), supports both root-velocity and pose-tracking command modes, and uses an MLP retargeting network plus a curriculum that starts with supervised initialization and continues with PPO. Experiments in IsaacGym on Unitree H1 and H1-2 compare against HumanPlus, ExBody, and OmniH2O on a held-out AMASS subset, reporting lower tracking errors; ablations isolate the dual-level architecture and the supervised-initialization curriculum. The paper also reports qualitative real-robot demonstrations on H1-2 and evaluates retargeting speed, accuracy, and smoothness against H2O and HumanPlus.","tokens_in":13113,"tokens_out":6478,"duration_ms":64474,"significance":"If the simulation results are taken at face value, the paper makes a useful empirical contribution: decoupling upper- and lower-body control improves upper-body pose tracking without a consistent penalty on lower-body tracking, the supervised-initialization curriculum stabilizes joint-angle tracking, and the MLP retargeter provides large speed gains with accuracy comparable to an optimization-based IK method. The internal ablations (Tables 4 and 5) are the strongest evidence because JAEGER, JAEGER Single, and JAEGER w.o. Init share the same reward and therefore isolate the architectural and curriculum effects. The retargeting comparison against H2O and HumanPlus is also a genuinely external benchmark. The main weaknesses are that the external SOTA comparison does not isolate the architecture from reward engineering, and the abstract's real-world superiority claim is not supported by quantitative hardware metrics. These gaps currently prevent the paper from fully supporting its strongest claims.","major_comments":[{"comment":"The headline comparison against HumanPlus, ExBody, and OmniH2O does not control for reward function or training curriculum. JAEGER uses a heavily hand-tuned reward (Table 1 and the appendix tables with weights spanning several orders of magnitude) and a multi-stage supervised-initialization curriculum, while the baselines were trained with their own reward designs and curricula and are not retrained under JAEGER's reward. Consequently the reported gaps (e.g., E_upper-j 0.1093 vs 0.2187 on H1 in Table 2) could reflect reward engineering rather than the dual-level controller itself. The internal ablations support the architectural claim, but they do not resolve this confound in the external comparison. The manuscript should either retrain the baselines with a matched reward or explicitly reframe the external comparison as a system-level comparison and soften the causal language. Section 6's own admission that the framework 'requires extensive reward engineering' makes this more than a stylistic point.","section":"Section 4.2, Tables 2 and 3"},{"comment":"The abstract claims superiority over state-of-the-art methods 'in both simulation and real environments,' and Section 4.2 repeats the simulation-side claim, but Section A.1 provides only deployment details (50 Hz policy inference, LCM communication, 500 Hz low-level interface) and qualitative video demonstrations. No quantitative hardware metrics are reported, such as tracking error, command-following success rate, or fall counts. The real-environment superiority claim is therefore currently unsupported. Please add quantitative real-world evaluations or revise the abstract and conclusion to claim only simulation superiority with qualitative real-world feasibility.","section":"Abstract and Section 4.2 vs. Section A.1"},{"comment":"The text states that JAEGER 'significantly outperforms' the baselines, but the only evidence is mean error with standard deviations across five seeds; no significance test is reported, and in some cells the intervals overlap (e.g., H1 root-linear-velocity in Table 2: 0.1504 +/- 0.0264 vs. HumanPlus 0.1696 +/- 0.0013). Given that the central claim is a comparative one, a paired or repeated-seed significance test, or at least a statement of whether the same evaluation trajectories and command samples are used across methods, would make the comparison more rigorous.","section":"Section 4.2 and Tables 2-3"}],"minor_comments":[{"comment":"The heading 'Regtargeting Network' contains a typo and should read 'Retargeting Network.'","section":"Section 3.1"},{"comment":"The captions for Tables 11 and 12 appear to be swapped: Table 11 is labeled 'Reward components and weights' but contains what look like PPO hyperparameters, while Table 12 is labeled 'Hyperparameters for RL Controller' but contains reward terms. Please check the numbering and make the captions match the content.","section":"Appendix C, Tables 11 and 12"},{"comment":"The sentence 'After training the models for both modes separately, we distill them into a single network via imitation learning' is inconsistent with the dual-level description elsewhere, which presents two independent policies. Clarify whether the deployed controller is one distilled network or two networks, and specify what is distilled (the two command modes, or the upper/lower policies).","section":"Section 3.3, final paragraph"},{"comment":"On H1-2, JAEGER's root-r (0.0398) and root-y (0.8051) are worse than OmniH2O's (0.0314 and 0.6540), so the claim that JAEGER achieves 'comparable accuracy in root tracking' should be made metric-specific rather than presented as a blanket statement.","section":"Section 4.2, Table 3"},{"comment":"For OmniH2O, the paper says reference keypoint positions are converted to joint angles 'to ensure a fair comparison,' but the conversion procedure is not described. Please specify the conversion method so readers can judge whether it favors any particular method.","section":"Section 4.1"},{"comment":"The paper does not announce a code or model release despite providing a project URL. Since the baselines are openly available and the evaluation is on a held-out AMASS subset, a public release of JAEGER's code and checkpoints would materially improve reproducibility and the value of the comparison.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The internal ablations are the most convincing part of the paper and suggest the architectural idea is sound. The main risk is that the SOTA-superiority claim is over-sold given the reward/curriculum confound in the external comparison and the absence of quantitative real-world metrics. I would support publication after the comparison is reframed or supplemented and the hardware claim is either quantified or trimmed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Ziluo and colleagues built JAEGER, a dual-level whole-body controller that splits upper and lower body into two policies. That is a real design idea, not just a hyperparameter change. The two policies share observations and rewards, and training uses a supervised initialization for the upper body with KL regularization to keep the lower body near a base policy. The internal ablations are the strongest part: replacing the dual-level with a single transformer (JAEGER Single) and removing supervised init (JAEGER w.o. Init) both hurt, and the gap is large especially for upper-body tracking. Since those ablations share JAEGER's reward and curriculum, the architecture and curriculum claims are supported. The retargeting MLP also looks fine: faster than H2O's IK and HumanPlus, and smoother on shoulder angles.\n\nThe soft spot is the comparison to HumanPlus, ExBody, and OmniH2O. Those baselines were trained with their own reward functions and curricula, and JAEGER's reward is heavily engineered (weights up to 10,000, multiple terms). So the apparent SOTA superiority in Tables 2 and 3 could come from reward engineering rather than the dual-level idea. The paper would be more convincing if the baselines were retrained with JAEGER's reward, or if a reward-matched single-policy baseline were used for all comparisons. JAEGER Single is close for one comparison, but it's their own architecture, not an external baseline.\n\nThe abstract claims superiority in real environments, but the real-world section is qualitative and the demos don't report error metrics. That overstates the evidence. No code is released, which hurts reproducibility.\n\nThe paper is honest about limitations: it requires extensive reward engineering and doesn't support keypoint commands. That counts in its favor. The citation pattern is fine; they acknowledge HOVER as concurrent and position themselves appropriately.\n\nWho is this for? Researchers in legged locomotion and humanoid teleoperation. It deserves a serious referee, and the revisions are tractable. I'd send it out, but I'd ask for reward-matched baselines, quantitative hardware data, and code or a clear release plan.","headline":"A genuinely different dual-level controller for humanoid whole-body control, with solid internal ablations, but the SOTA comparison is confounded by reward engineering and the real-world claim is unquantified.","tokens_in":13687,"tokens_out":1986,"would_cite":true,"duration_ms":20234,"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":"Splitting a humanoid's upper and lower body into two coordinated controllers cuts whole-body tracking error below single-controller baselines in simulation and on real hardware.","keywords":["whole-body control","humanoid robots","dual-level controller","multi-agent reinforcement learning","motion retargeting","curriculum learning","pose tracking","sim-to-real transfer"],"falsifier":"Retrain a single-controller baseline (a monolithic transformer with the same observations, actions, and reward as JAEGER) using JAEGER's supervised initialization, curriculum schedule, and domain randomization on the same motion dataset, then measure the same tracking metrics on the same evaluation set; if its errors match or beat JAEGER's, the dual-level architecture is not what drives the improvement.","tokens_in":12598,"feed_emoji":"🤖","tokens_out":9332,"duration_ms":82289,"temperature":0.7,"pith_summary":"The paper proposes JAEGER, a whole-body controller for adult-sized humanoid robots that splits control into two independent policies—one for the lower body, one for the upper body—instead of training one network to move all joints. The authors argue this separation reduces mutual interference between balance and expressive arm/head motion, shrinks the effective action space for each learner, and makes the system more fault-tolerant. They also introduce an MLP-based retargeting step that converts human motion-capture poses to humanoid joint targets faster and smoother than optimization-based inverse kinematics, and a curriculum that starts with supervised imitation and then refines with reinforcement learning. In simulation on two adult-sized platforms, JAEGER reports lower tracking errors than three open-source baselines in both root-velocity and pose-tracking modes, and the controller is demonstrated on real hardware. A sympathetic reader would take the core claim to be that decoupled, coordinated agents plus supervised initialization form a better inductive bias for whole-body humanoid control than a single monolithic policy.","feed_headline":"Two controllers beat one for whole-body humanoid control","feed_subtitle":"JAEGER splits upper and lower body policies, beating single-controller baselines in simulation and on real hardware.","key_machinery":"The load-bearing object is the dual-level controller, a Dec-POMDP with two agents: a lower-body policy implemented as a gated Transformer-XL taking 15 timesteps of observation history, and an upper-body policy implemented as a 3-layer MLP taking 3 timesteps. The curriculum's supervised initialization uses a per-step squared-error loss $\\mathcal{L}_{\\text{SL}}(\\pi_{\\text{upper}}) = \\|\\pi_{\\text{upper}}(x_{t,3}) - g^{\\text{upper}}_t\\|$ to make the upper body imitate the reference pose directly, and a KL term $\\mathcal{L}_{\\text{KL}}(\\pi_{\\text{lower}}) = D_{\\text{KL}}(\\pi_{\\text{lower}}(\\cdot|x_{t,15})\\,\\|\\,\\pi_{\\text{base}}(\\cdot|x_{t,15}))$ to keep the lower body near a pretrained base policy. These mechanisms, plus a three-layer MLP retargeting network trained on optimized human-to-humanoid pose pairs, carry the argument: the dual-level structure isolates interference, the supervised initialization prevents the upper body from becoming over-conservative and the lower body from forgetting balance, and the retargeting MLP supplies smooth, high-rate reference targets.","core_discovery":"JAEGER's central claim is that whole-body humanoid control should be formulated as a two-agent Dec-POMDP (decentralized partially observable Markov decision process): a Transformer-based lower-body policy handles balance, root velocity, and lower-limb pose, while a lightweight MLP upper-body policy independently tracks reference arm and torso angles. The two policies share observations and rewards, and their actions are concatenated into target joint positions for PD controllers. The paper reports that this dual-level controller, trained with supervised initialization followed by reinforcement learning, achieves significantly lower mean absolute tracking errors than single-controller baselines—for example on the 19-DoF platform in root-based mode, upper-body joint error of about $0.109$ rad versus at least $0.169$ rad for one baseline, and angular velocity error of $0.094$ rad/s versus $0.31$ rad/s for another. In pose-based mode, its upper- and lower-body joint errors are roughly 30% to 60% of the baselines' errors. The paper also claims the retargeting MLP runs above 1 kHz with smoother output than optimization-based inverse kinematics and lower error than directly copying human angles.","pith_inferences":["Editorial inference: the cleanest way to separate architecture from reward is to retrain a monolithic baseline with JAEGER's exact reward, curriculum, and randomization; if that baseline closes the gap, the dual-level structure is not the active ingredient.","Editorial inference: because the two policies are independent, a user could swap in a stronger upper-body policy or a different lower-body architecture without retraining the other half—an upgrade path the paper does not explicitly test.","Editorial inference: JAEGER omits keypoint-position commands; extending the upper-body policy to accept keypoint targets could combine the dual-level separation with the global-pose expressiveness of keypoint-based methods.","Editorial inference: the real-world stability reported here likely depends on the hand-tuned penalty terms in the reward; a systematic ablation of those penalties would show how much of the sim-to-real success is architectural rather than reward engineering."],"forward_implications":["If JAEGER's claims hold, a humanoid controller can serve both coarse root-velocity commands and fine pose-tracking commands in one system, without one capability degrading the other.","The dual-level design lets each half of the body use its own network architecture—a long-horizon transformer below, a cheap MLP above—so computational cost concentrates where balance decisions are hard.","Separating the controllers improves fault tolerance: if the upper-body policy fails or is disabled, the lower-body policy can keep the robot balanced and moving, because the two were trained independently.","The retargeting MLP's sub-millisecond per-frame cost makes real-time teleoperation practical, converting motion-capture input to joint targets at control rates above 1 kHz.","Supervised initialization before reinforcement learning appears to prevent the upper body from freezing into an over-conservative policy and keeps the lower body from drifting off its balance behavior during joint training."],"supporting_citations":[{"why":"Supplies the human motion database used for retargeting and for training and evaluating the controller.","marker":"[11]"},{"why":"Provides the HumanPlus baseline for comparison and its data-filtering procedure for high-quality retargeting.","marker":"[2]"},{"why":"Provides the OmniH2O baseline for pose-based comparison and the domain-randomization setup.","marker":"[3]"},{"why":"Provides the ExBody baseline for root-based comparison.","marker":"[6]"},{"why":"Supplies the two-stage optimization approach used to generate human-to-humanoid pose pairs for training the retargeting MLP.","marker":"[1]"},{"why":"Provides PPO, the base reinforcement-learning algorithm for the RL training stage.","marker":"[22]"},{"why":"Supplies the Gated Transformer-XL architecture used for the lower-body controller.","marker":"[21]"},{"why":"Provides the GPU-based simulation environment where all policies are trained and evaluated.","marker":"[23]"},{"why":"Provides the Dec-POMDP formulation that frames the two-agent control problem.","marker":"[18]"}],"fun_headline_variants":["Splitting humanoid control into two policies cuts tracking error","Decoupled policies outperform single whole-body controller","Upper and lower body policies beat joint controller","JAEGER's two-agent design improves humanoid tracking accuracy","Dual-policy setup cuts humanoid tracking errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise of the headline comparison is that the external baselines were evaluated under conditions that make JAEGER's advantage attributable to its dual-level architecture; because the baselines were not retrained with JAEGER's reward function and curriculum, the reported gap could be caused by reward engineering rather than by the controller split.","fun_headline_variants_meta":{"raw":{"variants":["Splitting humanoid control into two policies cuts tracking error","Decoupled policies outperform single whole-body controller","Upper and lower body policies beat joint controller","JAEGER's two-agent design improves humanoid tracking accuracy","Dual-policy setup cuts humanoid tracking errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000887,"raw_usage":{"total_tokens":3826,"prompt_tokens":940,"completion_tokens":2886,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2811}},"tokens_in":556,"tokens_out":2886,"duration_ms":20599,"temperature":1.0,"reasoning_tokens":2811,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:38:21.646729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain a single-controller baseline (a monolithic transformer with the same observations, actions, and reward as JAEGER) using JAEGER's supervised initialization, curriculum schedule, and domain randomization on the same motion dataset, then measure the same tracking metrics on the same evaluation set; if its errors match or beat JAEGER's, the dual-level architecture is not what drives the improvement.","supporting_citations":[{"cited_title":"Mahmood, N","cited_arxiv_id":null,"evidence_quote":"Supplies the human motion database used for retargeting and for training and evaluating the controller."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the OmniH2O baseline for pose-based comparison and the domain-randomization setup."},{"cited_title":"Cheng, Y","cited_arxiv_id":null,"evidence_quote":"Provides the ExBody baseline for root-based comparison."},{"cited_title":"Parisotto, F","cited_arxiv_id":null,"evidence_quote":"Supplies the Gated Transformer-XL architecture used for the lower-body controller."},{"cited_title":"Makoviychuk, L","cited_arxiv_id":null,"evidence_quote":"Provides the GPU-based simulation environment where all policies are trained and evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Dec-POMDP formulation that frames the two-agent control problem."}],"review_version":1}