Blind Dexterity: Whole-Body Humanoid Manipulation via Pure Proprioception

Aditya Bhatt; Oleg Kaidanov; Puze Liu; Jan Peters

cs.ROAug 30, 2026

Blind Dexterity: Whole-Body Humanoid Manipulation via Pure Proprioception

Authors: Aditya Bhatt, Oleg Kaidanov, Puze Liu, Jan Peters

Organizations: Intelligent Autonomous Systems Lab, TU Darmstadt · German Research Center for AI (DFKI) · Tongji University, Shanghai Research Institute for Autonomous Intelligent Systems · Robotics Institute Germany

Abstract

We present blind, whole-body manipulation skills on a Unitree G1 humanoid using only onboard proprioception, without cameras, markers, force-torque, or tactile sensors. Despite this minimal sensing, the trained policies exhibit surprising capability across qualitatively different tasks: push-resilient bipedal walking without IMU feedback, active soccer ball trapping with a foot, seeking and lifting a suitcase by its handle, and mounting a randomly positioned skateboard. We argue that these capabilities arise from a key underappreciated signal: the way the joint encoder readouts evolve under purposeful compliant contact, effectively forming a whole-body tactile channel. By generating contact-rich motions, the trained policies actively probe the environment; as a result, task-relevant object state (e.g., pose) becomes increasingly decodable from short proprioceptive histories. We expose this information using compact task-specific state estimators trained alongside, but fully separately from, the policies; their prediction errors decrease rapidly after informative contact. Our results indicate that joint encoder-based proprioception, combined with compliant actuation (now widely available on commercial robots and low-cost motors) is already a strong, practical substrate for whole-body dexterous manipulation and interactive perception, and therefore a natural foundation on which richer sensing can be layered.

Figures & tables

Fig. 1: Estimating inertial measurements from proprioception. In simulation, we command all three locomotion policies to stand in place, perturb the robot with an identical velocity impulse at $5s$ , and compare the predictions of the proprioceptive estimator against the ground-truth inertial measurements. The left column depicts a policy trained with access to IMU readings as well as the last commanded joint positions, the policy in the middle column does not receive IMU readings, while the third policy is additionally deprived of the previously commanded actions.

Fig. 2: Two football trapping episodes. The robot actively sweeps the region with one foot, makes contact, and traps the football underfoot.

Policy	Success (%)		Loc Error over Episode ( $\mathrm{cm}$ )
	Mean $\pm$ Std	Best	Mean $\pm$ Std	Best
Privileged	$99.8\pm 0.1$	$\mathbf{99.9}$	-	-
Distilled	$31.2\pm 33.3$	$67.7$	-	-
Blind -SE	$92.9\pm 1.2$	$94.3$	$9.51\pm 1.11$	$8.97$
Blind +SE	$86.5\pm 22.3$	$\mathbf{99.3}$	$13.16\pm 6.95$	$\mathbf{8.09}$

TABLE II: Policy and estimator performance on the football task. Aggregated over 5 random seeds (mean

\pm

std) and best seed.

Fig. 3: Active football localization. The robot must find and trap a randomly placed football ( red ) with its right foot. We visualize the estimator’s predicted ball position in transparent green . The blind policy exhibits active seeking ; it lifts the foot and wiggles it at the ankle while sweeping the area; upon making the first contact with the ball ( $0.73\mathrm{s}$ ) the estimate improves rapidly as the robot repeatedly taps the ball from above to stop it from rolling away, until it is firmly trapped with a localization error of less than $2\mathrm{cm}$ .

Fig. 4: Blind Skateboard Mounting. We show one blind +SE policy across three episodes with different initial skateboard placements ( red ); the predicted pose ( green ) is estimated from the same observations as the policy. At $0\mathrm{s}$ , the estimate is the mean over all possible initial poses. As in the football task, the policy first sweeps the area with one foot to “seek” the skateboard; the pose estimate improves upon first contact. A single touch only reveals the board’s presence at that location; its yaw and extent around the foot remain ambiguous. We then observe two additional interactive perception behaviors linked to further error reduction. First, touching the nose or tail —the inclined semicircular ends—produces a sharp drop in position error along the board’s major axis (Episode A, $0.27$ – $0.7\mathrm{s}$ ). If the nose is missed, touching the tail with the next foot (Episode B, $0.87\mathrm{s}$ ) yields the same effect. Second, dragging the board reliably induces motion along its major axis, sharply improving the yaw estimate. Across a successful episode, these behaviors generate informative proprioceptive signals that enable accurate skateboard pose estimation.

Policy	Success (%)		Pose Error over Episode
	Mean $\pm$ Std	Best	Mean $\pm$ Std	Best
Privileged	$28.1\pm 39.1$	$95.9$	-	-
Distilled	$8.2\pm 6.3$	$18.2$	-	-
Blind -SE	$88.2\pm 1.4$	$89.3$	$0.1101\pm 0.0106$	$0.1036$
Blind +SE	$89.7\pm 1.6$	$91.8$	$0.0981\pm 0.0073$	$0.0874$

TABLE III: Policy and estimator performance on the skateboard task. Aggregated over 5 random seeds (mean

\pm

std) and best seed.

Fig. 5: Blind Suitcase Lifting . We inspect the blind policy’s behavior over two episodes with different suitcase ( red ) sizes, positions, and table heights. The predicted suitcase pose and size ( green ) is estimated from the same proprioceptive observations that the policy receives. Until $0.43\mathrm{s}$ , the estimate is identical; it is the average of all initial suitcase poses during training. The policy first gently touches the table with the left “hand” ( $0.43-0.47\mathrm{s}$ ); the estimator instantly localizes the predicted suitcase vertically to lie on the true table surface. It then touches the suitcase with one or both hands from the sides, which strongly improves the complete pose estimate. Next, the robot waves its right hand near the top edge of the suitcase and searches repeatedly for the handle’s opening. Once the handle is “felt”, it commences insertion and lifting.

Config	Success (%)		Tipping (%)		SE Error (full)
	Mean $\pm$ Std	Best	Mean $\pm$ Std	Best	Mean $\pm$ Std
Privileged	$95.5\pm 0.3$	$\mathbf{95.9}$	$5.6\pm 1.4$	$\mathbf{3.6}$	-
Distilled	$34.6\pm 4.1$	$40.3$	$77.4\pm 3.3$	$72.3$	-
Blind -VS-SE	$83.5\pm 5.3$	$86.3$	$22.5\pm 2.8$	$22.3$	$0.3197\pm 0.0373$
Blind -VS+SE	$85.3\pm 1.9$	$87.9$	$23.9\pm 2.9$	$20.4$	$0.3358\pm 0.0304$
Blind +VS-SE	$88.3\pm 1.7$	$89.9$	$18.8\pm 3.1$	$19.7$	$0.3089\pm 0.1177$

TABLE IV: Suitcase evaluation results aggregated over seeds (mean

\pm

std).

Whole-Body Robotic Manipulation

Whole-Body Robotic Manipulation coordinates a robot’s arms, hands, body, and mobility to manipulate objects while maintaining balance and contact.

See all 88 papers

No Eyes, No Problem: What Does Touch Contribute to Blind Baoding Ball Manipulation?
Oct 16, 2026Ayush Deshmukh, Nalin Agarwal, Elle Miller +1In-Hand Manipulation Zero-Shot Robotic Manipulation
All authors are with the School of Informatics, University of Edinburgh, Edinburgh, UK.
Current as Touch: Proprioceptive Contact Feedback for Compliant Dexterous Manipulation
Jul 3, 2026Chenyang Ma, Yunchao Yao, Zhenyu Wei +3Robotic Control Compliant Robot Control
University of North Carolina at Chapel Hill
Learning Tactile-Aware Quadrupedal Loco-Manipulation Policies
Apr 29, 2026Pokuang Zhou, Yuhao Zhou, Quan Khanh Luu +7Legged Robotics RL for Robotics
Purdue University, West Lafayette, IN, USA. · Istituto Italiano di Tecnologia, Genoa, Italy. · Columbia University, New York, NY, USA.

Oct 16, 2026J/K move · Enter open · S save

No Eyes, No Problem: What Does Touch Contribute to Blind Baoding Ball Manipulation?

Ayush Deshmukh, Nalin Agarwal, Elle Miller, Sethu Vijayakumar

All authors are with the School of Informatics, University of Edinburgh, Edinburgh, UK.

In-Hand Manipulation Zero-Shot Robotic Manipulation Tactile Sensing for Robotic Manipulation Sim-to-Real Transfer Robot Manipulation Benchmarks

Open paper AskPDF arXiv

Humans manipulate objects within their hands through touch alone with remarkable ease, but for robots, achieving sustained real-world manipulation without vision remains an open challenge. Although reinforcement learning (RL) can generate sophisticated blind manipulation policies in simulation, transferring them to a real dexterous hand remains hindered by continuously changing contacts, motion-induced tactile information drift, and the gap between simulated and physical dynamics. We study this problem through Baoding ball rotation on the Shadow Dexterous Hand Lite, using only proprioception and binary tactile feedback as observations. We achieve successful zero-shot transfer through a combination of a sensor-aligned, coupling-aware embodiment in simulation, tactile calibration, and two novel episodic domain randomisations: slew-rate action limits and STuck-AT Taxels. On the physical hand, the policy achieves sustained blind Baoding rotation, completing

\textbf{225 consecutive

\mathbf{180^\circ}

rotations (112.5 full rotations) at

\sim0.36 ;\text{rot/s}

}

in a single continuous trial without any vision or object-state information. Furthermore, we evaluate the contribution of tactile sensing relative to proprioception alone and identify the hand regions where tactile information is most important. This is the strongest reported hardware result for Baoding ball manipulation in both rotation count and speed, irrespective of the sensing modality used. By effectively bridging the sim-to-real gap, our framework sets a new benchmark for sustained tactile-driven dexterous manipulation on physical hardware.

Abstract

Whole-Body Robotic Manipulation

Explore similar work

No Eyes, No Problem: What Does Touch Contribute to Blind Baoding Ball Manipulation?

Current as Touch: Proprioceptive Contact Feedback for Compliant Dexterous Manipulation

Learning Tactile-Aware Quadrupedal Loco-Manipulation Policies