PaperlayerView exact revision on arXiv

Source-linked AI summary

RoboSTD: Zero-Shot Single-to-Dual Transfer via Sagittal Mirroring for Bimanual Learning

Hongyu Wu, Chong Liu, Shaoxuan Xie, Shaqi Luo, Hanyu Feng, Shihan Wu, Guocai Yao

arXiv:2610.36685v1cs.RO

TL;DR

Bimanual learning depends on costly demonstrations, while reusing single-arm data is complicated by differences in action spaces, coordination, and arm-side preferences. RoboSTD mirrors single-arm demonstrations and uses LLM-guided reconstruction to create coordinated pseudo-bimanual supervision. In experiments, it achieved 2.11× the average number of collected cups over the single-arm baseline in the spatial generalization task.

  • Problem

    Collecting bimanual demonstrations is costly, and reusing single-arm data remains underexplored because of differences in action spaces, coordination requirements, and arm-side spatial preferences.

    Source: §I · p. 1

  • Method

    RoboSTD mirrors observations, proprioception, and actions to create contralateral pseudo-demonstrations, then uses LLM-guided reconstruction to coordinate skills as pseudo-bimanual supervision.

    Source: §I · p. 1

  • Results

    2.11× the average number of cups collected over the single-arm baseline was achieved in the spatial generalization task.

    Source: §I · p. 2

  • Takeaways & Limitations

    Experiments in simulation and real-world environments show that RoboSTD can provide effective pseudo-bimanual supervision, broaden workspace use, mitigate arm-side spatial preference, and improve long-horizon coordination.

    Source: §VI · p. 8

Abstract

from arXiv · show

Bimanual robot manipulation plays an important role in real-world robotic applications, but the high cost of collecting bimanual demonstrations limits the scalable development of bimanual policy learning. Meanwhile, existing single-arm manipulation data are difficult to directly transfer to bimanual systems due to the Single-to-Dual gap in action spaces, coordination requirements, and arm-side spatial preference. To address this issue, we propose RoboSTD, a zero-shot Single-to-Dual data construction method for bimanual robot learning. RoboSTD first applies sagittal-plane mirroring to observations, proprioception, and actions to construct executable contralateral pseudo-demonstrations. It then uses an LLM to generate structured spatio-temporal coordination constraints that guide the reconstruction of pseudo-bimanual supervision across workspace regions and task stages. We evaluate RoboSTD on the AgileX ALOHA platform in both RoboTwin simulation and real-world settings. Experiments demonstrate the effectiveness of RoboSTD for bimanual policy learning. Without requiring additional bimanual demonstrations, RoboSTD increases the average number of collected cups to 2.11x that of the single-arm baseline in the cup collection task. Project page: https://huahungy.github.io/RoboSTD/

I. INTRODUCTION

RoboSTD addresses the costly and underexplored reuse of single-arm demonstrations for bimanual learning, where differences in action spaces, coordination, and arm-side spatial preferences create a Single-to-Dual gap. It combines sagittal mirroring with LLM-guided coordination; experiments report effective supervision from fewer single-arm demonstrations and broader workspace use.

I. INTRODUCTION

  • Bimanual demonstrations are costly to collect, while reusing single-arm data remains underexplored because action spaces, coordination needs, and arm-side preferences differ.Collecting bimanual demonstrations requires recording both arms while maintaining spatial relationships, timing, and task-role coordination.Source: §I · p. 1
  • In 10,000 Open X-Embodiment trajectories, 35.7% of end-effector events were in the left workspace and 64.3% in the right; π0 showed a 9.6-point arm-side Bias Gap, versus 0.4 points for ACT.The authors suggest that the right-side preference in the data and pretrained policy may be reflected in arm-side performance.Source: §I · p. 1
  • RoboSTD uses sagittal-plane mirroring to create contralateral pseudo-demonstrations, then LLM-guided reconstruction to coordinate original and mirrored skills.The reconstruction uses arm assignment, precedence, and conflict constraints.Source: §I · pp. 1–2
  • 2.11× the average number of cups collected over the single-arm baseline was achieved in the spatial generalization task.In simulation, RoboSTD-generated supervision from 50 single-arm demonstrations per task reached a 66.5% average success rate, compared with 67.6% using 100 native bimanual demonstrations per task; long-horizon sandwich-making also improved in task success rate and average completion.Source: §I · p. 2

II. RELATED WORK

  • Prior work spans robot datasets and bimanual learning, demonstration augmentation, and transfer; AnyBimanual still requires additional bimanual demonstrations for adaptation.The paper positions RoboSTD as reusing single-arm demonstrations for bimanual supervision, unlike approaches that depend on additional bimanual demonstrations.Source: §II · p. 2

III. PROBLEM FORMULATION

The section frames Single-to-Dual transfer as constructing executable, coordinated bimanual supervision from single-arm trajectories. The challenge is that those trajectories lack the other arm’s state and actions, while cross-arm mappings and skill composition must preserve physical consistency and avoid spatial and temporal conflicts.

III. PROBLEM FORMULATION

  • Single-arm trajectories omit the contralateral arm’s state and action as well as annotations for subtask, arm role, and temporal stage.Bimanual supervision must represent both arms together with spatio-temporal coordination.Source: §III · pp. 2–3
  • Direct reuse is generally insufficient because single-arm observations cover only one side, while bimanual policies require coordinated states and actions for both arms.The action space expands from single-arm to dual-arm control.Source: §III · pp. 2–3
  • Cross-arm transfer must preserve consistency because differences in coordinate frames, joint axes, and mounting configurations can break correspondence between observations, proprioception, and actions.The goal is executable coordinated supervision without additional task-specific bimanual demonstrations.Source: §III · pp. 2–3
  • Combining valid original and mirrored skills alone does not guarantee coordination: arms can overlap in workspace or execute incompatible actions at the same stage.Long-horizon tasks require arm assignment and sequencing to avoid spatial conflicts and temporal ambiguity.Source: §III · p. 3

IV. METHOD

RoboSTD constructs pseudo-bimanual supervision from single-arm demonstrations, robot geometry, and task context. Its sagittal-plane mirroring jointly maps observations, proprioception, and actions to the contralateral arm while preserving physical correspondence.

IV. METHOD

  • RoboSTD combines single-arm demonstrations, robot geometry, and task context to construct pseudo-bimanual supervision without collecting additional task-specific bimanual demonstrations.Its pipeline first mirrors skills and then uses constraints to compose original and mirrored skills.Source: §IV · p. 3, Figure 3 · p. 4

A. Sagittal-Plane Mirroring

  • Sagittal-plane mirroring applies observation, state, and action mappings while exchanging arm identity, producing contralateral pseudo-demonstrations with preserved physical correspondence.The joint-state mapping is derived from joint-axis definitions and mounting configuration and applied consistently to corresponding states and actions.Source: A. Sagittal-Plane Mirroring · pp. 3–4, Figure 3 · p. 4
  • For mirrored end-effector poses, reflection maps position as ˜p = Sp and orientation as ˜R = SRS, with corresponding rotation-axis and angle-sign changes.The reflection matrix satisfies S = S^T = S^-1 and det(S) = -1.Source: A. Sagittal-Plane Mirroring · p. 3

B. LLM-Guided Spatio-Temporal Reconstruction

RoboSTD uses an LLM to infer structured coordination constraints from task context, then arranges original and mirrored skills into pseudo-bimanual supervision. The constraints specify arm assignments, precedence, and conflicts, while stage annotations condition the language-based policy.

B. LLM-Guided Spatio-Temporal Reconstruction

  • The LLM generates arm assignments, precedence relations, and conflict pairs from manipulation units and task context.The task context includes the instruction, object identities and locations, and workspace, reachability, and safety constraints.Source: B. LLM-Guided Spatio-Temporal… · p. 4
  • The rearrangement operator selects original or mirrored skills by arm assignment, orders them by precedence, and separates conflicting units in time.During a single-arm stage, the inactive arm holds its current configuration; the arm sequences are aligned on a common timeline.Source: B. LLM-Guided Spatio-Temporal… · p. 4
  • Each reconstructed stage receives a language annotation encoding its subtask, arm role, and temporal stage, which is composed with the task instruction for policy conditioning.This changes training supervision without changing the downstream policy architecture.Source: B. LLM-Guided Spatio-Temporal… · p. 4

C. Algorithm Summary

The algorithm derives mirroring transforms from robot geometry, generates coordination constraints, and rearranges original and mirrored skills into annotated pseudo-bimanual data. The experiments use AgileX ALOHA in RoboTwin simulation and real-world tasks, with supervision settings and metrics chosen for effectiveness, coordination, and arm-side spatial preference.

C. Algorithm Summary

  • Algorithm 1 derives the reflection matrix and joint mapping from robot geometry, mirrors each source trajectory, then constructs pseudo-bimanual data using LLM-generated constraints and stage annotations.The pipeline holds the inactive arm’s pose during single-arm stages.Source: C. Algorithm Summary · p. 4

A. Experimental Setup

  • Simulation uses a two-arm, 6-DoF AgileX ALOHA in RoboTwin 2.0 across 10 tasks; Single-Arm and RoboSTD use 50 source demonstrations per task, while Oracle uses 100 native bimanual demonstrations.Real-world evaluation covers five tasks, with 50 single-arm demonstrations for RoboSTD and 50 native bimanual demonstrations for Oracle per task.Source: A. Experimental Setup · p. 5
  • The experiments compare Single-Arm, RoboSTD, and Oracle supervision; high-coordination real-world tasks also include RoboSTD without LLM-generated coordination constraints.Evaluation reports task success rates, cup counts, sandwich completion level, Bias Gap, and Density Gap; a smaller positive Density Gap indicates more balanced left-right coverage.Source: A. Experimental Setup · pp. 5–6

B. Effectiveness of RoboSTD for Bimanual Learning

RoboSTD-generated supervision approaches native bimanual training in simulation and supports real-world transfer across tasks with different coordination demands. Real-world results also show an added benefit from LLM-guided coordination on high-coordination tasks.

B. Effectiveness of RoboSTD for Bimanual Learning

  • RoboSTD reaches 66.5% average success versus 39.1% for Single-Arm and 67.6% for Oracle across 10 RoboTwin tasks.The RoboSTD setting uses 50 single-arm demonstrations per task, versus 100 native bimanual demonstrations per task for Oracle.Source: B. Effectiveness of RoboSTD for… · p. 6

C. Bimanual Manipulation Across Coordination Levels

  • Cup Collection rises from 2.00 to 4.22 collected cups, above Oracle's 3.92; Sandwich completion reaches 74% versus Oracle's 72.5%.Without LLM-generated coordination constraints, RoboSTD reaches 2.76 collected cups and 63% sandwich completion.Source: Figure 5 · p. 6, C. Bimanual Manipulation Across… · p. 7
  • Across three real-world low-coordination tasks, mean success rises from 40.0% with Single-Arm to 77.3% with RoboSTD.The tasks cover rigid-object placement, deformable-object manipulation, and precision placement.Source: C. Bimanual Manipulation Across… · pp. 6–7
  • Mirroring and spatial rearrangement transfer skills, while LLM-guided reconstruction organizes them through arm assignment, precedence, and conflict constraints.For high-coordination tasks, geometric transfer alone remains below full RoboSTD.Source: C. Bimanual Manipulation Across… · p. 7

D. Mitigating Arm-Side Spatial Preference

RoboSTD reduces both trajectory-density imbalance and the performance gap between arms in selected RoboTwin tasks. The conclusion describes mirroring and LLM-guided reconstruction as complementary parts of its transfer framework.

D. Mitigating Arm-Side Spatial Preference

VI. CONCLUSION

  • The conclusion assigns mirroring the role of transferring skills to the contralateral workspace and LLM-guided reconstruction the role of organizing them for coordination.The stated constraints include arm assignment, precedence, and conflict.Source: §VI · p. 8

APPENDIX I DERIVATION OF MIRROR-CONSISTENT JOINT MAPPING

The appendix derives how sagittal reflection maps end-effector poses and joint states between robot arms. The joint-angle mapping depends on the arms' geometry and the transformation of their rotational axes.

  • The joint-space derivation assumes corresponding joints share actuation definitions and the zero-configuration structure is symmetric about the sagittal plane.Source: Appendix I · p. 10
  • Sagittal reflection transforms end-effector position and orientation consistently.The appendix describes the mirrored pose as a reflection of both position and orientation.Source: Appendix I · p. 10
  • A mirrored joint angle's sign depends on how its rotational axis relates to the reflected source axis.Reflection changes both the rotation axis and the sign of the rotation angle.Source: Appendix I · p. 10

A. Physically Mirrored Arms

For physical mirror copies, the arms' corresponding joint axes reflect across the sagittal plane, yielding a special-case joint-space mapping. For identical arms mounted symmetrically, the angle mapping depends on how each axis transforms under reflection.

A. Physically Mirrored Arms

  • For physically mirrored arms, the corresponding left-arm joint axis is the sagittal reflection of the right-arm axis, and the mapping is the special case Λ = −I.Source: A. Physically Mirrored Arms · p. 10

B. Pitch-, Yaw-, and Roll-like Joints in Isomorphic Arms

1) Pitch-like joints:

The joint-angle sign outcome follows how the joint axis transforms under sagittal reflection: pitch-like joints typically preserve sign, while yaw- and roll-like joints typically reverse it.

1) Pitch-like joints:

  • Pitch-like joints rotate about lateral axes approximately perpendicular to the sagittal plane, so reflection reverses those axes.Source: 1) Pitch-like joints · p. 10
  • Pitch-like joints therefore typically preserve their joint-angle sign under sagittal mirroring.Source: 1) Pitch-like joints · p. 11

2) Yaw- and roll-like joints:

C. Unified Representation

RoboSTD represents isomorphic-arm mirroring with a diagonal matrix of platform-specific joint signs. Those signs follow the reflected joint-axis geometry and robot mounting configuration, not joint names.

C. Unified Representation

Related papers

Loading 2610.36685v1…