Low-cost passive leader arms enable scalable data collection for imitation learning, but their lack of restoring forces leaves redundant joints unconstrained, yielding highly inconsistent joint-space trajectories. We address this with a low-cost passive joint module (an adjustable mechanical clutch and a 3D-printed spiral spring) that pulls each joint toward a user-defined neutral posture. This passive restoring force acts as a physical prior equivalent to joint-space L2 regularization (``Hardware-as-Regularization''), improving kinematic consistency with shorter completion time than an unregularized arm and preserved movement smoothness. Experiments show substantial variance reduction over a naive passive arm, trajectory consistency on par with an active virtual spring of the same actuator class, and substantially higher downstream success rates (e.g., 73.3% vs. 36.7% on Block Stacking); a stiffness sweep shows consistency degrading monotonically as stiffness decreases, confirming that mechanical stiffness behaves like a regularization weight.
Figures & tables
Figure 1: Overview of Hardware-as-Regularization. We introduce a purely mechanical joint module that consists only of 3D printable parts, screws, and bearings. By building a redundant teleoperation arm (e.g., GELLO) with this module containing a passive spring and an absolute encoder, we can mechanically regularize joint angles. This physical prior effectively solves the optimization problem minq21(q−q0)TK(q−q0) subject to the task-space constraint f(q)=xd , allowing users to collect highly consistent imitation learning data.
Metric
Naive GELLO
GELLO w/ v.s.
Proposed
Restoring Force
None
Active (Motor Torque)
Passive (Spiral Spring)
Data Consistency
Low (Null-space drift)
High
High (Physical Prior)
Latency
Zero
High (Control loop)
Zero (Continuous time)
System Stability
Passive (no energy injection)
Conditionally Stable
Passive (no energy injection)
Cost
∼ $200
> $700 (Current controllable motor)
∼ $100
Table 1: Comparison of Leader Arm Architectures for Data Collection
Figure 2: Exploded view of the proposed compliant joint module: a 3D-printed spiral spring for bidirectional compliance, a Hirth coupling secured via set screws for discrete neutral-position adjustment, and a coaxial magnetic encoder for joint-angle measurement.
Figure 3: Overview of the experimental setups. (a) Stiffness measurement with an analogue torque wrench on the proposed module. (b, c) The leader devices: a GELLO with virtual springs and the proposed passive module. (d) The simulation interface showing the current robot pose and the target goal pose (shadow). (e) The physical workspace for imitation learning with a 7-DoF arm, a fixed camera, and a wrist camera.
Figure 4: Results of the stiffness evaluation. (a) Torque versus displacement for the proposed passive spiral spring module (linear fit: 0.0330 Nm/rad). (b) Torque versus displacement for the active virtual spring, showing intermittent mechanical dead zones. (c) Commanded and measured current versus displacement for the virtual spring (linear electrical response: 29.45 mA/rad).
Figure 6
Figure 6: Comparison of individual operator performance across three kinematic consistency metrics: (a) joint standard deviation, (b) Pairwise DTW, and (c) trajectory smoothness (SPARC). Note that User 1 is an expert, while Users 2–5 are novices. The proposed module consistently minimizes null-space drift (a, b) for each user regardless of skill level, while preserving natural smoothness (c).
Task
Metric
Naive GELLO
GELLO w/ v.s.
Proposed
Pick & Place
Pairwise DTW [deg] (Data)
23.23 ± 11.01
11.71 ± 5.51
8.32 ± 3.14
SPARC [-] (Data)
-2.87 ± 0.23
-3.23 ± 0.23
-2.97 ± 0.25
Success Rate (95% CI) [%]
53.3 (36.1, 69.8)
96.7 (83.4, 99.4)
96.7 (83.4, 99.4)
Stack
Pairwise DTW [deg] (Data)
16.35 ± 6.94
10.23 ± 3.99
8.24 ± 2.99
SPARC [-] (Data)
-2.87 ± 0.21
-3.89 ± 0.30
-3.19 ± 0.33
Success Rate (95% CI) [%]
36.7 (21.9, 54.5)
66.7 (48.8, 80.7)
73.3 (55.6, 85.8)
Table 3: Comparison of collected-data quality and downstream ACT success rates (95% Wilson score intervals; n=30 rollouts per condition). Naive GELLO trails the proposed module significantly on both tasks (two-sided Fisher exact test), while proposed vs. virtual spring is not significant on either task (see text).
Imitation learning for robotic manipulation relies on large sets of human demonstration trajectories, which are often noisy and temporally irregular due to variable operator speed, intermittent pauses, and inconsistent action density. A common preprocessing strategy is time-uniform downsampling to shorten sequences, but it cannot effectively remove speed-induced non-uniformity or redundant pauses. This mismatch degrades data quality and hinders policy learning. To address this issue, we propose Information-Standardized Trajectory Resampling (ISR), an offline preprocessing method for effective imitation learning. ISR resamples each trajectory by enforcing approximately equal information distance between adjacent points. Specifically, we map trajectories onto an information-modulated Riemannian manifold and perform geodesic-equidistant parameterization. We construct an information-intensity field from velocity and acceleration norms: the velocity term removes small-motion redundancy, while the acceleration term preserves high-curvature and fine-manipulation phases. We evaluate ISR on three real-world manipulation tasks with mainstream imitation learning policies. Compared with the baseline time-uniform 3x downsampling, ISR improves task success rates by about 25%, remains robust across datasets collected from different operators, and reduces both dataset size and training cost. The code and videos are publicly available at https://d-robotics-ai-lab.github.io/isr.page.
Licheng Yang, Lingfeng Qian, Fei Zheng +4
State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences · D-Robotics
Robot manipulation policies trained through imitation learning commonly rely on costly demonstrations recorded in one task direction and at one temporal resolution. We propose Reversal and Stride-based Trajectory Augmentation (RESTA), which exploits time-scaling data augmentation to improve data efficiency through trajectory reversal and temporal resampling. Trajectory-reversal augmentation reverses and relabels demonstrations between paired tasks with opposite objectives, reusing recorded rollouts without additional environment interaction. Because contact-rich manipulation is not exactly reversible, we treat reversal as an approximate paired-task prior rather than a physical symmetry. Temporal-stride augmentation varies observation and action sampling strides and conditions the policy on the resulting offsets, enabling fixed or randomized temporal schedules. Experiments across 16 paired bimanual manipulation tasks show that trajectory reversal is particularly effective in the low-data regime and that temporal-stride augmentation improves success under randomized temporal schedules. With 25 and 50 independently collected demonstrations per task, RESTA achieves 78.4% and 85.2% average success, respectively, matching or exceeding baselines trained with twice as many demonstrations (77.3% and 85.4%). These results show that RESTA can improve data efficiency in robot imitation learning.
Yunpeng Jiang, Jianshu Hu, Han Fang +3
Global College, Shanghai Jiao Tong University, Shanghai, China · Shanghai Institute of Physical Intelligence and Robotics (PAIR), Shanghai, China · Duke Kunshan University, Jiangsu, China
Retargeting human kinematic reference motion onto a robot's morphology remains a formidable challenge. Existing methods often produce physical inconsistencies, such as foot sliding, self-collisions, or dynamically infeasible motions, which hinder downstream imitation learning. We propose a bilevel optimization framework that jointly adapts reference motions to a robot's morphology while training a tracking policy using reinforcement learning. To make the optimization tractable, we derive an approximate gradient for the upper-level loss. Our framework requires only a sparse set of semantic rigid-body correspondences and eliminates the need for manual tuning by identifying optimal values for a parameterization expressive enough to preserve characteristic motion across different embodiments. Moreover, by integrating retargeting directly with physics simulation, we produce physically plausible motions that facilitate robust imitation learning. We validate our method in simulation and on hardware, demonstrating challenging motions for morphologies that differ significantly from a human, including retargeting onto a quadruped.