Conventional robot arms can be extremely heavy compared with the everyday objects they manipulate, creating a significant inertia mismatch at contact. Even a small accidental bump from a heavy arm can knock a light object out of place. Modern learning methods have nevertheless demonstrated a wide range of successful manipulation behaviors on these platforms, suggesting that even extreme inertia mismatch can be overcome with sufficient data. However, it remains unclear whether high inertia imposes a hidden cost on learning, potentially worsening scaling with data or lowering the performance ceiling of modern robot learning methods. We investigate this question through controlled simulation experiments that vary robot and object masses. We find that reinforcement learning succeeds across a broad range of mass combinations, while behavior cloning learns more reliable policies when a lighter robot manipulates a heavier object, and more data narrows the resulting gap without closing it. Behavior cloning likewise degrades on a simulated robot arm as its motors reflect more inertia through their gearing, even though its weight stays the same. Finally, because new hardware is rarely an option, we explore ways to lower contact inertia on the arm one already has, either by choosing a lower-inertia workspace before collecting data or by steering the demonstrator toward lower-inertia solutions of the task. In both simulated cases, the lower-inertia demonstrations are imitated more reliably. Project website: https://younghyopark.me/inertia-mismatch/
Figures & tables
Fig. 1: Robot arms can significantly outweigh the everyday objects they manipulate. A natural question arises: can modern robot learning fully compensate for this mismatch given sufficient data, or does it impose a hidden cost on learning reliable manipulation?
Fig. 2: Robots far outweigh objects in public demonstrations. Mean masses; dashed line: y=x .
Fig. 3: A mass-scaled spring and damper drive a point-mass robot that pushes a T. A circular robot with mass mr pushes an object with mass mo toward the outlined goal. The policy commands a position xd , which the robot tracks through a virtual spring kp and damper kd . These gains scale with mr , so that every robot has the same unsaturated free-space response for a given natural frequency ωn and damping ratio ζ .
Fig. 4: Mass-scaled gains give each controller setting one free-space response at every robot mass. Each panel shows the analytic steady-state response to a 1 Hz sinusoidal target: columns vary natural frequency ωn ( 5 and 40 rad/s), and rows vary damping ratio ζ ( 0.5 and 2 ).
Fig. 5: RL succeeds across the whole mass grid. Arrival success, averaged over controller settings and finish times (85–94%).
Fig. 6: BC, unlike RL, strongly prefers a lighter robot manipulating a heavier object. Each policy is a Transformer trained on 200 demonstrations: from four steps of robot and object observations, it predicts five absolute planar position targets and executes one before replanning ( H=4 , P=5 , E=1 ). All panels share mass axes and a color scale, saturating below 50%, for arrival success at each run’s best closed-loop evaluation. Panels (a–e) average the controller and finish-time settings done for every variant; (f–g) split the baseline by finish time. The table-friction expert uses a higher amax , so the lightest robot can slide the heaviest T.
Fig. 7: More data narrows the gap but does not close it. A heavy robot needs 16 times the demonstrations to match a light robot with 100, and even 6,400 leave it below a light robot with 200. Curves show the mass-ratio extremes of Fig. 6(a) over 18 settings with 95% intervals; policies train for 50k updates, and success averages the last five closed-loop evaluations. The expert’s success differs far less than BC’s between the extremes (90% and 86%).
Fig. 8: Training error does not follow the mass preference. Mean training MSE before each run’s best closed-loop evaluation, averaged over the settings in Fig. 6(a) .
Fig. 9: BC success falls as a robot reflects more motor inertia, at the same weight. For each robot: its task, the RL expert’s success, and BC’s success with 200 demonstrations (three data draws per cell), over the reflected-inertia scale s and the object’s mass; s=1 is the unscaled robot. The arm’s expert reaches at least 98% arrival success in every cell; the hand’s expert falls the same way as BC, but by less. The two color scales differ, and (b) leaves out s=0.1 and the 1 kg cube, whose corner weakens the trend without reversing it.
Fig. 10: Elevating the workspace lowers the inertia the same arm presents along planar push directions. Each glyph plots the directional effective mass at one tip position over planar push directions: its radius and color both grow with mass, with identical scales in (b) and (c). The smaller glyphs in (c) identify a placement where the robot presents less inertia to planar contact, so such a map can reveal a lower-inertia workspace before collecting demonstrations or training a policy. The dashed circle is a radius key, gray crosses mark positions unreachable with the vertical tool, and the star marks the robot base. Red and blue sheets in (a) follow the mapped footprints at z=0 and 0.40 m; the gray grid marks the robot’s base plane.
Fig. 11: Moving the task improves imitation without replacing the robot. Higher rollout success in (b) comes without a consistent change in validation action error in (c), where the two placements overlap across experts: the lower-inertia placement produces a more reliable policy without fitting its held-out demonstrations consistently more accurately. MLP policies train on 200 demonstrations to predict absolute XY tip targets from histories of tip position and object pose. Panels (b–e) show settings picked for large gains, with the largest, H=8,P=4,E=1 , in (b–c); Fig. 13 shows all 36. Panel (a) shows the baseline (left, red) and lower-inertia (right, blue) placements from the same view; sheets follow the regions in Fig. 10 , gray grids mark z=0 , and green marks the goals. Curves average three experts and three data draws per expert; shading is sample SD across expert means. MSE uses each condition’s own validation data; success axes emphasize late training.
Fig. 12: Workspace placement lowers inertia along the demonstrated motions. Curves show the median and interquartile range over 1,800 training trajectories per placement, aligned by normalized elapsed time. Mass is the directional effective mass of Eq. ( 9 ) along measured tip XY velocity, omitting near-stationary samples.
Fig. 13: All 36 settings gain from the lower-inertia placement. Dots: settings, averaged over expert seeds and data draws; marker and bar: mean ± SD at 150k updates; pp: percentage points.
Fig. 14: The placement gain persists at lower policy rates. Gap between the two placements’ best settings at 50k updates.
Fig. 15: Lower-inertia demonstrations teach more reliable joint-position policies. BC trains on 2,000 demonstrations from the low- or high-inertia expert, with the same robot and position controller. Higher success in (b) accompanies larger action MSE in (c). From histories of joint positions and cube pose, plus the goal pose, policies predict a chunk of absolute joint-position targets, converted from the experts’ desired torques in the demonstrations, and execute one target before replanning. Panels (b–e) show the multi-seed settings that gain most, with the largest, H=5,P=10 , in (b–c); Fig. 17 shows all 16 at 50k. Panel (a) overlays recorded expert frames at similar cube poses, showing distinct arm configurations; one cube pose is shown for clarity, and the green goal pose is illustrative, repositioned and reoriented to show the task. Shading is sample SD across learner seeds. Training MSE uses a log axis; success axes emphasize late training.
Fig. 16: Two RL experts solve the same task with distinct contact inertia. Median and interquartile range across 2,000 paired successful demonstrations per condition, from one expert each, aligned by normalized elapsed time. The reward and this plot evaluate m along the tool–cube contact normal, during approach and contact, with mref=2 kg, βlow=4 s -1 , and βhigh=1 s -1 .
Fig. 17: Most settings gain from lower-inertia demonstrations. Gain at 50k updates over 16 settings; marker and bar: mean ± SD.
Robotic manipulation robustness often founders on the physics gap between simplified simulations and the resistance-laden real world. In this work, we emphasize that physical realism in articulated interaction is an important ingredient for robust policy learning. We present Real-IKEA, a dataset and simulation framework designed with physical accuracy as a first-class goal. Real-IKEA provides 1,079 articulated asset configurations, derived from 83 authentic IKEA handles and knobs processed through a meticulous six-step physical workflow. For contact-geometry accuracy, we introduce a bidirectional surface-deviation metric to quantify collision meshes. For dynamics realism, we establish resistance-calibrated configurations that vary damping and friction. Crucially, we demonstrate through a Reinforcement Learning (RL) policy that high-fidelity assets enable the discovery of robust "hooking" and "levering" strategies that prioritize mechanical advantage over fragile friction-pulling. Together, these results position Real-IKEA as a critical benchmark for developing manipulation policies capable of human-level robustness in articulated object tasks.
Kunqi Xu, Zhenhao Huang, Siyuan Luo +2
Human-centered Robotic Lab, National University of Singapore · School of EECS, Peking University
Position controllers have become the dominant interface for executing learned manipulation policies. Yet a critical design decision remains understudied: how should we choose controller gains for policy learning? The conventional wisdom is to select gains based on desired task compliance or stiffness. However, this logic breaks down when controllers are paired with state-conditioned policies: effective stiffness emerges from the interplay between learned reactions and control dynamics, not from gains alone. We argue that gain selection should instead be guided by learnability: how amenable different gain settings are to the learning algorithm in use. In this work, we systematically investigate how position controller gains affect three core components of modern robot learning pipelines: behavior cloning, reinforcement learning from scratch, and sim-to-real transfer. Through extensive experiments across multiple tasks and robot embodiments, we find that: (1) behavior cloning benefits from compliant and overdamped gain regimes, (2) reinforcement learning can succeed across all gain regimes given compatible hyperparameter tuning, and (3) sim-to-real transfer is harmed by stiff and overdamped gain regimes. These findings reveal that optimal gain selection depends not on the desired task behavior, but on the learning paradigm employed. Project website: https://younghyopark.me/tune-to-learn
Robot-learning studies often compare position and velocity actions as though each label identified a single experimental treatment, yet the label alone does not determine what the robot executes. Before affecting robot motion, bounded policy output passes through filtering, physical scaling, reference construction, feedback, timing, and effort limits. We study how these omitted properties affect attribution under one shared learner and training budget. Across three contact-rich Factory tasks (PegInsert, GearMesh, and NutThread), we compare five joint-space interfaces, with ten independently trained policies per condition. We first remain within the same position-feedback law and vary the executed interface: substituting only the physical action scale as the directly modified component changes sample-median evaluation success by 0.645-1.000 across tasks, while removing an exponential moving average produces task-dependent changes. We then move to a cross-law comparison and choose a velocity scale that exactly matches the position interface's local raw-torque command gain. This calibration raises the sample median on GearMesh and NutThread, with no median change on PegInsert. Yet clear residual sample-median gaps remain on PegInsert and GearMesh, while the NutThread interval includes zero. Thus, even this explicit equality leaves feedback and held-action dynamics unmatched and does not establish position-velocity trained-outcome equivalence. Action-interface comparisons should therefore report the executed pathway and state the physical quantity they match.
Arshia Sangwan, Zhongyang Zhang
New York University · University of California San Diego