{"id":"be155bab-753c-4d33-b2b5-cbaa319c0517","arxiv_id":"2509.09364","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A $6,380, 110 cm, 14.5 kg 3D-printed humanoid using off-the-shelf backdrivable actuators demonstrates walking, jumping, fall mitigation, and stand-up, with open-source design files.","lead":"AGILOped is a small, open-source humanoid robot that can walk, jump, brace for falls, and stand up again, built mostly from 3D-printed parts and off-the-shelf motors. The full design is published so other labs can build a capable research robot for about $6,400 instead of buying a closed commercial platform.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (1) kinematic constraint is unverified under load: backlash/flexure in the double 4-bar parallelograms would break the passive ankle and foot-lock, corrupting joint estimates and balance control, and undermining the open-source high-performance claim.","rationale":"The reader's weakest assumption is exactly the one I consider most load-bearing. Eq. (1) is not a minor detail; it is the conceptual centerpiece that enables a cheap, light, 10-actuator design with a passive ankle. If it fails, the robot loses its foot orientation reference, the Jacobian torque mapping is wrong, and the whole control architecture—including the ZMP/CoM balance feedback shown in Fig. 6—is built on a false premise. This is a correctness risk, not just a parameter-tuning issue. It is also empirically unaddressed: the experiments are qualitative and would be consistent with a robot whose ankle passively gives way at the moment of impact. The paper's reliance on NimbRo-OP2X's years of RoboCup use is not a substitute for AGILOped-specific stiffness/backlash data, since that robot is heavier, larger, and uses different joint construction. Other weaknesses (lack of repeated trials, no walking speed) are important but are evidence-quality issues; the kinematic constraint is a structural assumption that, if wrong, would invalidate the central claim outright. Thus I agree with the reader's CONDITIONAL verdict and see no reason to change it.","tokens_in":10783,"tokens_out":9387,"duration_ms":99415,"concrete_test":"Attach optical motion-capture markers to the thigh, shank, and foot of one leg. Run the robot through a jump cycle and a deliberate fall, recording marker positions at 200 Hz. Compare the measured q_h, q_k, q_a to the values inferred from the motor encoders and Eq. (1). Repeat 10 times, then repeat after 100 falls. If the RMS error exceeds ~2° or if a backlash hysteresis loop greater than ~1° appears, the kinematic constraint is not reliable under dynamic loading, and the central claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's key design is the double 4-bar parallelogram leg (Eq. (1)) that locks foot orientation to the torso via a passive ankle and reduces actuator count. This is central to both the low cost (10 actuators for 12 joints) and the dynamic capability. The control system infers joint angles from motor encoders and applies the constant Jacobian J (Eq. (2)) to map torques. This entire chain implicitly assumes the links are rigid and backlash-free. The actual links are 3D-printed plastics plus manually-threaded aluminum rods with screwed-in universal joints; any play or flex in these links under dynamic loads (jumping, fall impacts) makes Eq. (1) approximate at best, and the passive ankle/foot-lock fails. The paper provides no measurements of joint-angle error, backlash, or link stiffness, and the long-term durability claim is borrowed from NimbRo-OP2X, which is a different, larger, bolted design. If the constraint is not tightly satisfied, the IMU-kinematic state estimate and the balance controller inherit errors, and the demonstrated walking/get-up robustness may not transfer to end-user builds, directly weakening the 'accessible high-performance' claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents AGILOped, an open-source 110 cm, 14.5 kg humanoid robot built from off-the-shelf components and 3D-printed structural parts, with a total cost of about $6,380. The key mechanical contribution is a double 4-bar parallelogram leg that couples hip, knee, and ankle motions with only four actuators per leg, giving a passive ankle whose foot orientation is nominally locked to the torso. The authors describe the actuator electronics, low-level impedance control, a CPG-based walking controller, and a keyframe-based jump, fall-mitigation, and get-up system. Experimental sections show forward walking on artificial grass, vertical jumps of 'about 10 cm', and a sequence of a backwards fall followed by standing up. The central claim is that AGILOped closes the gap between high-performance and accessible humanoid platforms for research.","tokens_in":11095,"tokens_out":3551,"duration_ms":43497,"significance":"If the claims hold, AGILOped would be a valuable contribution: it is one of the most affordable adult-sized open humanoid platforms with dynamic capabilities, and the open CAD/BOM/software lowers the entry barrier for research groups. The design choice of a parallel-kinematic leg with only ten actuators is elegant and potentially lowers both cost and leg inertia. The paper gives cost breakdown, dimensions, electronics schematics, and a clear kinematic model, and reports real hardware demonstrations rather than simulations only. The reported 3D-printed construction with off-the-shelf actuators is in line with the authors' prior NimbRo work, lending some credibility to the platform's practicability. However, the quantitative evidence is thin, and the load-bearing kinematic constraint is not validated under dynamic load. If those gaps are addressed, this could become a useful reference platform paper.","major_comments":[{"comment":"The kinematic constraint Eq. (1) is load-bearing: the double 4-bar parallelograms with 3D-printed links and manually-threaded aluminum rods with screwed-in universal joints are assumed rigid and backlash-free, yielding the constant Jacobian in Eq. (2). The paper provides no measurement of joint-angle error, backlash, or deflection under dynamic loading. If the links flex or the universal joints have play, the passive ankle and foot-orientation locking fail, corrupting the joint estimates used by the state estimator and balance controller. I ask for a concrete validation: e.g., compare motor-encoder-derived joint angles with visual marker measurements during walking/jumping, or report a stiffness/backlash test under representative loads.","section":"Section III-B, Eq. (1)-(2)"},{"comment":"The jump experiment reports 'about 10 cm' without repeated trials, error bars, or a success criterion. The statement that the robot 'jumps slightly backwards' suggests a systematic CoM/control issue that is not analyzed. To support the high-performance claim, please report the number of trials, mean and variance of jump height, the landing success rate, and ideally ground-reaction-force data synchronized with a motion-capture height measurement.","section":"Section V-B, Jumping"},{"comment":"The walking demonstration is qualitative: Fig. 6 shows CoM/ZMP traces and force plots, but no walking speed, step length, step frequency, number of consecutive steps, or trial count is given. Since the robot is claimed to 'walk forward' and the gait is a core capability, at least walking speed and step metrics are needed to assess performance and reproducibility against other platforms.","section":"Section V-A, Walking"},{"comment":"The durability/robustness claim rests largely on prior NimbRo-OP2(X) experience, not on AGILOped-specific testing. The text states that AGILOped was 'pushed from any side', but only a backwards fall is shown; other fall directions and impact cases are not reported. As fall robustness is central to the 'mitigation and getting-up' demonstration, please provide AGILOped-specific evidence: number of falls from each direction, inspection results for structural parts and parallel-rod joints, and any repairs or failures encountered.","section":"Section III-B, V-C, Falling Mitigation"}],"minor_comments":[{"comment":"Typo: 'RaspBerry Pi' should be 'Raspberry Pi'. Also '1 GBRAM' should have a space.","section":"Table II"},{"comment":"The min() operation is unclear for vector torque: is it element-wise clipping of each component to tau_max, or a scalar limit on the torque vector? Please clarify notation.","section":"Eq. (3)"},{"comment":"The footnotes e/f on Unitree G1 are hard to interpret ('non-programmable' and 'for developers' are ambiguous). The table header abbreviations 'Max. HFE Tor.' and 'Max. KFE Tor.' should be expanded for readability.","section":"Table I"},{"comment":"Spelling inconsistency: 'Nvidia' and 'NVidia' are both used. Also, 'varioShore TPU' is not defined; a brief material explanation or citation would help readers unfamiliar with this product.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The core platform idea is sound and the open-source contribution is valuable. The main technical gap is validation of the parallel-kinematic constraint under load; without it, the 'accessible high-performance' claim is not fully supported. The paper is borderline between a systems-description paper and an experimental validation paper; adding the requested quantitative data would make it acceptable. Scope fit for a conference like ARM is reasonable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"AGILOped is a real piece of work: a child-sized humanoid you can build for $6,380 from off-the-shelf parts and a 3D printer, and the authors show it walking, jumping, falling, and getting up. The hardware description is detailed enough to reproduce, and the open-source CAD, control stack, and a $30 AHRS module are genuine contributions to a subfield where most platforms are closed or unaffordable. I believe the central claim — that AGILOped gives mid-sized research groups access to a dynamic humanoid — is supported.\n\nThe soft spots are real but not disqualifying. The evaluation is mostly qualitative. There are no repeated trials, error bars, walking speed, step metrics, or a durability count on this specific robot. The jump height is reported as 'about 10 cm' and the paper itself admits it's not optimized and that the CoM model is off. That honesty is good, but it also means the dynamic capability is demonstrated rather than characterized.\n\nThe more specific concern, which I share with the stress-test note, is Eq. (1). The whole leg design rests on the assumption that the double 4-bar parallelograms are rigid enough that q_h = q_th, q_k = q_sh - q_th, q_a = -q_sh holds under load. The links are 3D-printed parts with manually threaded aluminum rods and screwed-in universal joints; any play or flex breaks the passive ankle and foot-lock, and the joint reconstruction from motor encoders inherits the error. The paper gives no measurement of backlash, stiffness, or kinematic accuracy. The long-term durability claim is borrowed from NimbRo-OP2(X), which is a different, larger design. This is a gap in characterization rather than evidence that the mechanism doesn't work — the robot demonstrably does work — but it matters for end-user reproduction. If someone else's print tolerances are worse, the same control stack may not be as robust.\n\nOne minor point: the cost table compares against a small set of platforms, and some prices are estimates, so 'most affordable' is a careful claim but not a rigorous market survey.\n\nWho is this for? Labs that want to get into humanoid locomotion research without a six-figure budget, and anyone who values reproducible hardware. It deserves a serious referee; a revision should add quantitative walking metrics, repeated jump trials, and ideally some backlash/flex measurements or a repeatability test. I'd engage with it.","headline":"A genuinely open, low-cost child-sized humanoid that walks, jumps, and gets up, with the kinematic chain's rigidity asserted rather than measured.","tokens_in":11559,"tokens_out":2727,"would_cite":true,"duration_ms":25998,"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":"AGILOped: a $6,380 open-source humanoid robot that walks, jumps, mitigates falls, and stands up, built from 3D-printed parts and off-the-shelf actuators.","keywords":["open-source humanoid","quasi-direct-drive actuator","3D-printed robot","parallel kinematics","passive ankle","bipedal walking","fall mitigation","low-cost robotics"],"falsifier":"Measure the actual passive ankle angle during walking or jumping using external motion capture or high-speed video, and compare it with the value predicted from the thigh and shank encoders via Eq. (1). A systematic error or growing backlash with repeated falls would show the kinematic model breaks. Alternatively, run repeated fall tests until structural failure to test the durability claim.","tokens_in":10681,"feed_emoji":"🤖","tokens_out":2756,"duration_ms":34159,"temperature":0.7,"pith_summary":"This paper presents AGILOped, a 110 cm, 14.5 kg open-source humanoid robot that costs about $6,380 and uses only a 3D printer, off-the-shelf backdrivable actuators, and standard electronics. The authors claim this platform closes the gap between high dynamic performance and accessibility, showing experiments in forward walking, vertical jumping of about 10 cm, controlled falling with impact mitigation, and autonomous standing up after a fall. If these results hold, any research lab with a 3D printer and modest budget could build and modify a humanoid capable of behaviors usually reserved for expensive or closed platforms. The design relies on a parallel parallelogram leg mechanism with a passive ankle, keeping inertia low and cost down, while the open CAD and software invite community customization.","feed_headline":"Open-source humanoid walks, jumps, and stands up for $6,380","feed_subtitle":"A 110 cm, 14.5 kg 3D-printed robot with off-the-shelf actuators brings dynamic behaviors to any research lab.","key_machinery":"The double 4-bar parallelogram linkage in each leg, which couples the thigh and shank actuators to the hip, knee, and passive ankle joints. It reduces the number of actuators needed for a 5-joint leg, creates a fast knee with torque shared between thigh and shank, and keeps the foot orientation horizontally constrained to the torso, lowering reflected inertia. The backdrivable quasi-direct-drive actuators (MyActuator RMD X6-40) with integrated controllers provide proprioceptive torque feedback and impedance control, and the whole system runs on a Raspberry Pi with a ROS-based software stack.","core_discovery":"The central claim is that a carefully designed, mostly 3D-printed humanoid using off-the-shelf quasi-direct-drive actuators can achieve dynamic whole-body behaviors at a fraction of the cost of existing research platforms. AGILOped has 10 actuators controlling 12 joints: each leg uses a double 4-bar parallelogram that lets two hip actuators drive hip pitch, knee pitch, and a passive ankle pitch, with the foot orientation locked to the torso. This kinematic coupling (q_h = q_th, q_k = q_sh - q_th, q_a = -q_sh) reduces actuator count and leg inertia, and the authors show experimentally that the resulting robot can walk forward with feedback control, jump, survive falls with compliant TPU parts","pith_inferences":["If the kinematic coupling holds precisely, AGILOped becomes a natural testbed for studying how a passive ankle affects balance and gait, since the ankle is not directly actuated—a comparison against an ankle-actuated variant would isolate that contribution.","The reported 10 cm jump height is limited by the lack of a dedicated force controller; implementing one could raise jump height, a testable extension of the paper's keyframe-based approach.","The no-backlash and long-term durability claims are extrapolated from prior NimbRo robots rather than AGILOped-specific measurements; direct measurement of joint backlash and fatigue under repeated falls would validate or refute this transfer.","The open-source release may enable crowdsourced hardware iterations, such as different feet, arms, or actuator replacements, creating a community benchmark for low-cost dynamic humanoids."],"forward_implications":["If the platform works as reported, research groups can build a dynamic humanoid for roughly $6,380 plus 3D-printing time, enabling reproducible experiments and hardware modifications across labs.","The passive-ankle parallelogram design offers a template for reducing leg inertia and actuator count in other open humanoid designs, potentially speeding up whole-body control and reduced-order model research.","The demonstrated fall mitigation and get-up routine suggest that 3D-printed structures with TPU compliance can endure repeated impacts, extending the practical lifetime of low-cost humanoids.","The open CAD and software stack could accelerate the adoption of learning-based and model-based control approaches by providing a fully transparent hardware platform.","The low-cost AHRS solution and simple electronics lower the software and sensing barrier for new robotics research groups entering the field."],"fun_headline_variants":["AGILOped: open-source robot that jumps and gets up","Jumping humanoid for $6,380—now open-source","Affordable humanoid: walks, jumps, and falls safely","110-cm open-source humanoid with agile moves","Humanoid research platform at a fraction of the cost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The kinematic constraint of the double 4-bar parallelogram—that the passive ankle angle follows q_a = -q_sh with no play or flex—must hold under dynamic loads; any backlash or deformation in the 3D-printed links or manually assembled rods would invalidate the leg model, balance estimation, and gait control.","fun_headline_variants_meta":{"raw":{"variants":["AGILOped: open-source robot that jumps and gets up","Jumping humanoid for $6,380—now open-source","Affordable humanoid: walks, jumps, and falls safely","110-cm open-source humanoid with agile moves","Humanoid research platform at a fraction of the cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000276,"raw_usage":{"total_tokens":1447,"prompt_tokens":670,"completion_tokens":777,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":414,"completion_tokens_details":{"reasoning_tokens":693}},"tokens_in":414,"tokens_out":777,"duration_ms":8661,"temperature":1.0,"reasoning_tokens":693,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:11:55.280401+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual passive ankle angle during walking or jumping using external motion capture or high-speed video, and compare it with the value predicted from the thigh and shank encoders via Eq. (1). A systematic error or growing backlash with repeated falls would show the kinematic model breaks. Alternatively, run repeated fall tests until structural failure to test the durability claim.","supporting_citations":[],"review_version":1}