← Paper Digest

Hidden Variables Behind Real-Robot Data Quality: Lessons from Hundreds of Thousands of Hours

Maniformer Real-World Manipulation Team, Che Tsang Calvin, Xingyu Chen, Bowen Jiang, Naixiang Gao, Xiongfeng Cai, Enlin Gu, Ruiying Li, Peize Li, Mianrui Zhu, Haomiao Zhang, Xin He, Minhui Yu, Jun Zhao, Zihao Zhu, Yang Chen, Jingmei Jin, Keru Zhou, Chen Wang, Chang Zhou, Zhiwen Hou, Di Yang, Wenjun Sun, Fan Yang, Liyan Zhang, Yi Lu, Haojun Ma, Jun Zhu, Pengxin Zha, Zhaosheng Wang, Xin Zhao, Xindong He, Chiming Liu, Jianheng Song, Jiayao Ma, Maoqing Yao, Guanghui Ren, Qian Wang, Wei Shan, Yang Yang, Li Liu, Wenhao Wang

November 2026

A wheeled humanoid sorts soft parcels from a bin with both arms.

TL;DR A large real-robot data-collection team measures how embodiment, teleoperation interface and calibration shape demonstrations, using 25 robot models, controlled teleoperation trials and real-robot π0.5 policies. Driving body parts in parallel cuts task time by 43-53%, and policies keep about 90% of the extra path length in detour-filled demonstrations.

What they do

A team that has logged 207,030 hours of real-robot demonstrations across 20,253 task records asks how the collection system itself shapes the data: which poses the robot can reach, how smoothly an operator can drive it, and whether the recorded signals explain the actions. Each factor is tested with kinematic analysis, controlled teleoperation trials, or real-robot evaluation of fine-tuned π0.5 policies.

How it works

  • Embodiment. URDFs of 25 dual-arm robots; IK tests 64 tool-axis directions at each of 10,075 positions, measuring orientation coverage instead of reach alone.
  • Teleoperation control. Hand controllers with mode switching are compared with M-to-N control (headset drives base and head, IK the waist, controllers the arms), and force with position control on contact tasks, 30 trials per configuration.
  • Data quality. π0.5 policies with delta joint actions are trained on demonstrations that differ in whether the camera saw the decision cue, in injected joint-zero offsets, and in the share of detour-filled trajectories.

Results

  • Humanoid and wheeled-lift arms reach 55% and 53% of a shoulder slice but cover only 41% and 33% of tool orientations, against 91% and 75% for table-mounted arms. Letting the shoulder translate 15 cm enlarges the reachable cross-section 3.4-5.5x.
  • M-to-N control cuts median completion time by 53% (floor pick) and 43% (wiping). Force control succeeds in every trial on door opening, wiping and box carrying, against 80-87% for position control, with 55-71% lower peak forces.
  • Active-search demonstrations raise correct arm choice from 50% to 100%. Keeping 25 clean episodes among offset ones gives 66.7% success versus 26.7% without; a miscalibrated inference robot fails all 30 trials. Policies keep about 90% of the extra wrist path from detour demonstrations, and 80% detour data lengthens execution by 29%.

Limitations

The embodiment study uses URDFs and discretized IK, ignoring collisions and dynamics. Teleoperation tests cover a few tasks, operators and devices. Data-quality tests use specific tasks, one policy and one action representation, with artificially injected offsets and sometimes a single training run.

Videos

A pedestal dual-arm robot and a humanoid move their arms side by side.

Figures

From Figure 1: Operators collect demonstrations through a range of human interfaces (left); the data (center; 200K+ c...
From Figure 1: Operators collect demonstrations through a range of human interfaces (left); the data (center; 200K+ collection hours across 20K+ tasks) train policies across robot embodiments (right). The study covers three factors that shape these data: embodiment, teleoperation control and data quality.
From Figure 5: Input-channel mapping. Left (1-to-N): a single input channel drives all body parts, which take turns v...
From Figure 5: Input-channel mapping. Left (1-to-N): a single input channel drives all body parts, which take turns via mode switching. Right (M-to-N): three channels drive five parts, the headset for base and head, IK for the waist and controllers for arms and grippers, so all parts can move at once.
From Figure 12: Demonstration setups. LOS-known: the operator sees the block, the robot camera does not, and the arm ...
From Figure 12: Demonstration setups. LOS-known: the operator sees the block, the robot camera does not, and the arm is chosen before seeing. BLOS-known: neither operator nor robot camera sees the block initially, and the arm is still chosen before seeing. BLOS-search: the robot camera sees the block before the arm is chosen.
From Figure 14: Training mixtures for the joint-zero-offset experiment, each 100 episodes in four groups of 25. A: al...
From Figure 14: Training mixtures for the joint-zero-offset experiment, each 100 episodes in four groups of 25. A: all clean. A': 25 clean plus three groups with biased offsets. B: no clean data, zero-mean offsets. B': no clean data, biased offsets. The table lists per-joint offsets in degrees; bars show real-robot pick-and-place success.
Abstract

As real-robot data collection scales, collection-system design shapes both operator efficiency and the behaviors learned by robot policies. Workspace limitations, control-interface friction, and latency can introduce pauses, detours, and corrective motions into otherwise successful demonstrations. We investigate these effects across embodiment, teleoperation control, and data consistency, combining a kinematic study of 25 dual-arm robots, controlled teleoperation experiments, and real-robot evaluation of fine-tuned $π_{0.5}$ policies. The embodiment study shows that positional reachability overstates orientation-aware task coverage and that modest shoulder motion can substantially expand the effective workspace. Teleoperation experiments show that parallelizing body-part control reduces task completion time by 43-53%, while compliant control improves success and reduces peak contact forces on constrained and contact-rich tasks. Policy-learning experiments show that demonstrations must capture the visual information underlying action choices: active-search demonstrations improve arm-selection accuracy from chance (50%) to 100%. Calibration experiments reveal that mixed-calibration training data remain usable in our tested configurations when an inference-aligned subset is retained, whereas inference-time miscalibration causes all 30 trials to fail. Policies also retain approximately 90% of the extra path length in detour-contaminated demonstrations, showing how inefficient operator motions persist in learned behavior. Together, these findings connect collection-system design to demonstration quality and downstream policy execution. They motivate evaluating effective workspace, control efficiency, observation completeness, calibration alignment, and motion quality before scaling collection, and addressing system-induced compensations at their source.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.