AthenaZero: A low-inertia, bimanual robot for dynamic manipulation
Authors: Andrew S. Morgan, Gregory Xie, Capprin Bass, Rachel Thomasson, Chunpeng Wang, Erfan Shahriari, Harrison Busa, Oluwaseun Araromi, Joseph Aronov, Michael Burgess, Velin D. Dimitrov, Matthew A. Estrada, Faris Hamdi, Samuel Kendig, Taeyoon Lee, Jose Oscar Mur-Miranda, Emma Sommers, Paul Titchener, Margaret Wang, Achu Wilson, Mark Yeatman, Osman Dogan Yirmibesoglu, Alfred A. Rizzi, Annan Mozeika, Nicolas Rojas, Lael Odhner
Organizations: RAI Institute, Cambridge, Massachusetts, 02142, USA
AthenaZero is a bimanual manipulator designed to minimize inertia without compromising control authority. By utilizing quasi-direct drive actuation and transmission remotization techniques, the system achieves an effective endpoint mass comparable to that of a human---about an order of magnitude less than conventional robot manipulators. This characteristic, combined with its inherent torque transparency, makes AthenaZero exceptionally well-suited for dynamic manipulation. We describe the methodology} that led to this design and demonstrate the robot's capabilities on three baseball-inspired tasks: throwing, catching, and batting, which showcase complex interactions on human-comparable timescales where milliseconds matter. AthenaZero was capable of throwing at speeds in excess of 30 m/s, with catching and batting at speeds in excess of 14 m/s over a short 7.3 m distance. Batting practice and a game of catch were subsequently performed in robot-to-robot and human-to-robot variations, showcasing the efficacy and adaptability of our system in tasks that require high acceleration.
Figures & tables
Figure 1: AthenaZero throwing a baseball. AthenaZero ’s low-inertia design lends itself to highly dynamic, athletic tasks such as throwing a baseball in excess of 25 m/s.
Figure 2: Motor-level and joint-level architecture of AthenaZero . Nomenclature: Torso (T), Right (R), Left (L), Joint (J), and Motor (M) . Two torso motors (TM1 and TM2) connect directly to the TJ1 torso joint in parallel via a belt. Shoulder motors RM1-RM3 (LM1-LM3) are axially coincident, combining towards RJ1-RJ3 (LJ1-LJ3) and acting as a spherical shoulder joint. Elbow junction actuators RM4-RM5 (LM4-LM5) through a belt transmission are axially perpendicular, mimicking a universal joint RJ4-RJ5 (LJ4-LJ5) at the elbow. Wrist motors RM6-RM7 (LM6-LM7) are differentially coupled, symbolically representing a near-universal joint RJ6-RJ7 (LJ6-LJ7) at the wrist junction. Translational offset between the two axes is 6 mm. A total of 3 motors for each hand are located within the torso, with Bowden transmissions that remote actuation out to three fingers via tendons.
Group
Joint
Motor Diameter
Gear Ratio
Peak Torque
Peak Velocity
Actuator Inertia*
Torso
TJ1
2 x 95 mm †
10.48:1
378 Nm
32 rad/s
5.49e-2 kg m 2
Shoulder
(R/L)J1/2/3
95 mm
5:1
90 Nm
66 rad/s
6.25e-3 kg m 2
Elbow
(R/L)J4
95 mm
5:1
90 Nm
66 rad/s
6.25e-3 kg m 2
(R/L)J5
76 mm
5:1
20 Nm
89 rad/s
1.28e-3 kg m 2
Wrist
(R/L)J6
38 mm ||38 mm ‡
∼ 10.6:1
∼ 16.1 Nm
∼ 100 rad/s
4.09e-4 kg m 2
(R/L)J7
38 mm|| 38 mm ‡
∼ 7.5:1
∼ 11.4 Nm
∼ 142 rad/s
2.28e-4 kg m 2
Table 1: AthenaZero joint specifications. † TJ1 is actuated by two 95 mm motors coupled in parallel with a belt reduction. ‡ The wrist is a differentially coupled parallel mechanism. Wrist specifications are taken at the neutral position and due to the properties of the linkage, is non-linear. * Actuator Inertia signifies the total reflected rotor inertia of that motor/gearbox combination.
Figure 3: Remoting actuation via transmissions reduces effective mass. Shoulder and Elbow: The majority of the arm’s mass is contained via the combination of motors RM1-RM5 (equivalently LM1-LM5). Shoulder motors are located at the joint, but elbow flexion leverages a belt transmission. Wrist: The wrist is a 2 DoF parallel mechanism which remotes the motors to the elbow junction. Hand: The hand is a 6 DoF, 3-tendon underactuated mechanism that interdigitates. Tactile sensors are placed inside of the fingertips and palm for contact sensing (denoted in blue).
Figure 4: Effective mass analysis of AthenaZero and off-the-shelf manipulators. We compare the effective mass of a human arm, AthenaZero —with a locked torso—and a set of commercially-available manipulators (Barrett WAM, Franka Research 3, UR5e, and KUKA LBR iiwa 14). A. We compute effective mass in a neutral configuration for an operational point at the wrist across a planar sweep of directions. Solid lines report total effective mass (including reflected inertia of the actuators) whereas dashed lines report the effective mass of the links alone. Numerical values for the maximum effective mass across swept directions for each arm are shown. B. To promote whole-body contact, we compare the effective mass along all body structures, computed in the direction normal to the surface at each point. QDD manipulators ( AthenaZero / WAM) elicit a similar effective mass on their upper arms as the high-geared manipulators do at the end-effector. C. Effective mass is configuration dependent. We report the distribution of mean effective mass at the wrist across 2000 random configurations. Box limits indicate the 25th and 75th percentiles, the central line represents the median, and whiskers span the range of non-outlier data, extending to the most extreme values within 1.5 times the interquartile range of the box limits. Points falling beyond the whiskers are plotted individually as outliers.
Figure 5: Pendulum impact test. A. A 1 kg pendulum impact test on AthenaZero results in a relatively large end-effector displacement of 115.3 mm and a positive post-contact pendulum velocity. B. A 1 kg pendulum impact test on on the FR3 results in a much smaller end-effector displacement (13.7 mm) and a negative pendulum velocity (bouncing effect). Error bars indicate 95% confidence intervals calculated from the standard error of the mean for five independent trials per robot.
Figure 6: Robot stiffness evaluation. Resultant positional offset of AthenaZero when commanded to 30 random joint configurations in unloaded and loaded conditions. In the loaded condition, the 1.8 kg mass was affixed to the end effector. Box limits indicate the 25th and 75th percentiles, the central line represents the median, and whiskers span the range of non-outlier data, extending to the most extreme values within 1.5 times the interquartile range of the box limits. Points falling beyond the whiskers are plotted individually as outliers.
Skill
Maximum Velocity
Accurate Velocity
Throwing
30.8 m/s
21.4 m/s
Maximum Velocity @ 7.3 m
Reaction Time
Extrapolated Velocity @ 18.4 m
Catching
18.3 m/s
0.398 s
46.1 m/s
Batting
13.9 m/s
0.525 s
35.0 m/s
Table 2: Overview of throwing, catching, and batting capabilities. We showcase the maximum velocities of each baseball task, in addition to the equivalent catching and batting velocity according to reaction time if the two robots were placed on a standard pitcher’s mound and in the batter’s box (18.4 m).
Figure 7: Baseball-inspired tasks to showcase dynamic manipulation. A. Throwing a baseball at 13.4 m/s. B. Catching a baseball at 13.4 m/s @ 7.3 m (0.544 s reaction time). C. Batting a tennis ball at 13.4 m/s. @ 7.3 m (0.544 s reaction time).
Figure 8: Cooperative variations of playing a game of catch and participating in batting practice. A. Two robots playing a game of catch with one another. B. A robot and a human playing a game of catch. C. One robot throwing batting practice to another robot. D. A human throwing batting practice to a robot.
Joint Angle [rad]
AthenaZero
WAM
FR3
UR5e
iiwa14
q1
-0.1396
0
0
0
0
q2
0.1396
-1.5708
-1.5708
-3.14159
-1.5708
q3
0.5236
0
0
1.5708
0
q4
1.5708
1.5708
-1.5708
3.14159
-1.5708
q5
1.5708
0
0
-1.5708
0
q6
0
0
3.14159
0
0
Table 3: Neutral robot joint configurations. The specified joint angles place compared robots in similar poses for effective mass analysis.
Figure S1: Layered control architecture of AthenaZero . High-level planners generate motion and control parameters that are evaluated within a deterministic 1 kHz real-time loop to produce joint- or task-space reference signals. These references propagate through the control and runtime layers and are ultimately translated into high-frequency actuator commands over EtherCAT.
Figure S2: Dynamometer characterization of output torque in linear regime. Leveraging a dynamometer, we characterize the torque production capabilities of a QDD actuator. During a locked rotor test, the system swepdf different commanded torques and measures the resultant torque through a load cell. The data are fit via least squares, and a highly linear model representation with an R2 =0.9973 is noted for 2,741 samples.
Dynamic manipulation is a critical capability for robots operating in complex and dynamic environments, where robots must interact with objects that are moving or require rapid adjustments. However, learning models for dynamic manipulation tasks face two major challenges: (1) the combinatorial complexity of dynamic scenarios leads to substantial data requirements, and (2) rapid variations in dynamics require real-time and accurate policy execution. In this paper, we propose DynamicManip to address these challenges through an efficient data augmentation pipeline and a low-latency imitation policy. We first propose a static-to-dynamic augmentation pipeline that synthesizes diverse dynamic manipulation demonstrations from a single static demonstration. Second, we introduce a dynamic-aware adaptive policy that adjusts its inference frequency according to task dynamics, enabling responsive and effective dynamic manipulation. Third, we build a dynamic manipulation benchmark, which includes diverse dynamic tasks with an automatic evaluation system for scalable and consistent assessment. Extensive experiments in both simulation and the real world demonstrate that DynamicManip not only provides significant improvements in data efficiency but also achieves better performance in dynamic manipulation tasks, with a mean success rate 18.4 percentage points higher and policy-query latency 32.9% lower.
Haoran Liao, Pengyue Wang, Shuoyu Chen +10
1Sun Yat-sen University · 4Jiangxing Intelligence · 2Stanford University +1
Multi-stream robot manipulation policies achieve unparalleled sample efficiency and generalization by modeling actions relative to environmental reference frames. However, existing approaches typically assume these frames to be strictly exogenous. This causal assumption collapses in dynamic settings, such as when a single robot arm manipulates a moving object or when two arms coordinate, where each arm effectively becomes part of the dynamic environment of the other. We propose DynaMAC, a lightweight, policy-agnostic framework that resolves this causal limitation while preserving the sample efficiency, computational speed, and flexibility of multi-stream policies, DynaMAC treats the opposite arm as a dynamic task parameter, thereby providing a unified formulation for dynamic manipulation and bimanual coordination without requiring an explicit leader-follower relationship. To rigorously evaluate these capabilities, we introduce DynaBench, a novel benchmark for robot manipulation in dynamic environments. Across both dynamic environments and bimanual manipulation tasks, DynaMAC outperforms leading probabilistic and generative baselines by over 35 percentage points while requiring 20 times fewer samples. Crucially, DynaMAC generalizes zero-shot from static demonstrations to dynamic environments, substantially simplifying data collection and establishing an elegant bridge toward human-robot collaboration.
Jan Ole von Hartz, Abhinav Valada, Joschka Boedecker
Department of Computer Science, University of Freiburg, Germany.
We present a bimanual mobile manipulator built on the open-source XLeRobot with integrated onboard compute for less than $1300. Key contributions include: (1) optimized mechanical design maximizing stiffness-to-weight ratio, (2) a Tri-Bus power topology isolating compute from motor-induced voltage transients, and (3) embedded autonomy using NVIDIA Jetson Orin Nano for untethered operation. The platform enables teleoperation, autonomous SLAM navigation, and vision-based manipulation without external dependencies, providing a low-cost alternative for research and education in robotics and robot learning.