acceptodds
Under review as a conference paper at ICLR 2027

DexArena: A Unified Benchmark for Long-Horizon Dexterous Manipulation.

Abstract

Long-horizon dexterous manipulation remains a fundamental challenge for robot learning, requiring the composition of multiple skills, precise bimanual coordination, generalizable visual understanding, and robust skill transitions. Yet existing manipulation benchmarks largely emphasize isolated skills or support only a limited range of end-effectors, making these capabilities difficult to evaluate systematically. We introduce **DexArena**, a benchmark designed to evaluate learning and generalization in long-horizon bimanual dexterous manipulation. DexArena features **10 long-horizon tasks and 30 stage-level tasks** spanning diverse household and industrial scenarios. Each long-horizon task is composed of explicitly defined stages with physically grounded success conditions, enabling reliable evaluation of both intermediate progress and end-to-end task completion. A unified benchmark interface and verified human-teleoperated demonstrations enable testing different policy learning algorithms, with standardized simulation configurations, scene randomization, trajectory replay, and task-level evaluation. Through extensive baseline experiments, we identify substantial challenges in precise manipulation, generalization to scene variation, and skill composition. DexArena establishes a challenging and reproducible benchmark for developing robot learning methods that move toward generalizable long-horizon dexterous behavior.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Dexterous Manipulation, Benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_BoKQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a benchmark designed to evaluate dexterous manipulation policies on a wide range of tasks on long horizon bimanual manipulation. They include a multitude of tasks and demonstrations. On the dataset side, the primary delta over prior work (DexJoco) is the number of tasks provided and the framework for composing stage-level tasks into longer horizon tasks.

Critical strengths
  • Enabling task composition provides a useful interface for future work to study how the interplay of different tasks affect policy performance.
Critical weaknesses
  • Despite using a dexterous hand, most of the tasks studied don't really require any dexterity and could be tackled with a parallel jaw gripper. There are a few tasks requiring tool use (aspirate liquid), but for the most part, the tasks are long-horizon composition of simple primitives. The introduction motivates dexterous hands with “regrasp, twist objects, and manipulate tools” but the benchmark does not test any of this. Given that the whole point of using a dexterous embodiment is to tackle tasks which cannot be completed with a parallel jaw gripper, this seems like a major oversight.

  • While the paper states bimanual manipulation multiple times, nearly all of the tasks are for just one of the left/right hands. There is only one task(!) in the appendix that uses both hands, making bimanual a significant overclaim. Similarly, they say that they "support four end-effector types" but all data is collected with the Sharpa Hand and is kinematically retargeted, without any downstream verification of whether the demonstrations are successful.

Questions for authors
  • If any IL baseline is tested with the parallel-jaw gripper embodiment, would it be able to succeed? Ideally tested without retargeted data, but I understand that it may not be possible to collect fresh parallel-jaw data for a task. If tasks can actually be completed without dexterous hands, that weakens the usefulness of the benchmark
  • What percentage of the demonstrations are successful after retargeting? Does the benchmark provide a meaningful way to study cross-embodiment learning?
1 comment

We sincerely thank the reviewer for the careful reading and constructive feedback. Below, we clarify the role of dexterity in the current task suite and the scope of the bimanual and cross-embodiment evaluation.

W1 & Q1. Dexterity requirements and parallel-jaw baselines

We agree that some tasks in the current suite may also be solvable with a parallel-jaw gripper. However, the goal of DexArena is not to claim that every task is exclusively solvable by a dexterous hand, but to provide a benchmark for learning manipulation policies under a dexterous-hand embodiment and high DoF action space.

The current submission does not include an IL baseline evaluated with the parallel-jaw gripper. At the same time, we acknowledge that a direct comparison between dexterous-hand and parallel-jaw policies would be valuable for characterizing where dexterous embodiment provides additional capability. We also see value in expanding the task suite with more skills involving regrasping, in-hand rotation, and dexterous tool use.

W2 & Q2. Cross-embodiment and bimanual scope

We would like to clarify that “retargeting” in our teleoperation pipeline refers to mapping the operator’s hand motion to the robot hand during data collection. The released demonstrations themselves contain Sharpa Hand trajectories. Among the 4,112 accepted Sharpa episodes, 3,222 (78%) successfully reproduce the task under open-loop action replay, and the released 3,000 demonstrations are selected from these replay-successful episodes.

This replay evaluation validates the collected Sharpa trajectories, but does not constitute quantitative validation of transferring these demonstrations to the other supported embodiments. We currently do not report policy or demonstration-transfer results for the Wuji Hand, XHand, or gripper.

Similarly, OpenPackage is the only stage that requires simultaneous use of both hands, while most other stages are executed by a single arm within the dual-arm environment. We therefore view broader cross-embodiment evaluation and richer genuinely bimanual tasks as valuable directions for extending the benchmark.

Reviewer_iAoROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces DexArena, a simulation benchmark for long-horizon dexterous manipulation. It consists of 10 long-horizon tasks, each broken into 2–5 stages, giving 30 stage-level tasks that can be evaluated individually or chained together. Each stage has a physically grounded success condition, and the benchmark includes 3,000 VR-teleoperated demonstrations collected with a dexterous hand. The authors evaluate RL, imitation learning, and VLA baselines on the individual stages, under visual scene changes, and on the full long-horizon tasks. While many individual stages are learnable, the baselines struggle with precise contact-rich skills and rarely complete the full long-horizon tasks. The authors position DexArena as a common testbed for measuring dexterous skill learning, visual generalization, and skill composition.

Critical strengths
  1. Collecting high-quality teleoperation data for dexterous manipulation is challenging, and the authors put significant effort into their infrastructure for this. Demonstrations are collected with VR and data gloves, verified through state restoration, and filtered by open-loop replay. This dataset and pipeline could be useful for future practitioners.

  2. The long-horizon task suite is challenging for existing methods, covering skills beyond standard pick-and-place such as unscrewing, inserting, pipetting, and typing. The benchmark can in principle separate failures in individual skills from failures in composing them into full tasks. Current baselines perform poorly on the long-horizon tasks as well as on a number of the individual skills, which highlights a clear gap and gives a target for future methods.

Critical weaknesses
  1. The scope of the benchmark is limited. For a benchmark paper, there is relatively little variation within tasks to test targeted generalization: the generalization evaluation only swaps between three kitchen backgrounds, without varying object instances, geometry, or physical properties. The task suite itself is also small, with 10 long-horizon tasks. Finally, although the benchmark supports four end-effectors, all released data and experiments use a single dexterous hand, so the cross-embodiment support is not demonstrated.

  2. Compounding the limited scope of the benchmark, the results that are included receive little analysis, so there is not much a reader can take away from the experiments. The stage-level and generalization results are presented largely as tables with little discussion: there is no analysis of why methods differ (e.g., ACT at 58% vs. Diffusion Policy at 39% average stage success with the same inputs), which scene changes cause the large drops under visual variation, or what makes certain stages difficult. The authors compute detailed contact statistics for each task but relate them only to open-loop replay success, not to policy performance. Finally, there is no failure analysis on the long-horizon tasks, even though the stage-aligned design is meant to enable it.

Questions for authors
  1. Many stage-level tasks have near-zero success across all methods (e.g., InsertRail, DispenseLiquid, PickDebris), with each policy trained on roughly 90 demonstrations of a single skill. What is the main bottleneck on these tasks: the amount of data, the quality of the teleoperated demonstrations, or the intrinsic difficulty of the skill? Evaluating at least one method on the hardest stages with additional demonstrations or with privileged object state would help separate these factors.

  2. Every policy, including GR00T, is trained on a single skill's demonstrations. How would a single multi-task policy trained on all 3,000 demonstrations perform? This seems like the setting where a pretrained VLA would be expected to benefit most, and it would test whether skills transfer across the benchmark's tasks.

  3. On the long-horizon tasks, how far do rollouts get before failing? For example, ACT succeeds on the two Package opening stages at 44% and 88% individually, which would suggest roughly 39% end-to-end success if stages failed independently. Yet it completes the full task only 4% of the time. Reporting per-stage progress within long-horizon rollouts would help clarify whether failures come from individually weak stages, compounding errors, or some mismatch during stage transitions.

2 comments

We sincerely thank the reviewer for the careful reading and constructive feedback. Below, we clarify the current evaluation of challenging stage-level tasks, multi-task training, and long-horizon execution, and discuss how broader task variation and generalization could strengthen the benchmark.

W2 & Q1. Bottlenecks on difficult stage-level tasks

We agree that the current experiments do not disentangle whether the failures on the hardest stages are primarily caused by limited data, demonstration quality, or the intrinsic difficulty of the underlying contact-rich skills. Our replay filtering verifies that the released demonstrations can be successfully reproduced under open-loop replay, but this does not establish that the amount or diversity of the data is sufficient for policy learning. Controlled experiments that vary demonstration scale or provide privileged state information would be useful for separating these factors, and we consider this an important direction for future evaluation.

Q2. Multi-task policy training

For the stage-level evaluation, the baselines are trained separately for individual stages. For the long-horizon GR00T evaluation, we additionally train one policy per long-horizon task using the pooled demonstrations of its constituent stages. However, we do not evaluate a single policy trained jointly on all 3,000 demonstrations across the complete benchmark.

We agree that this broader multi-task setting would be valuable, particularly for pretrained VLA models, since it could reveal whether representations and manipulation skills transfer across different task families. Extending the benchmark toward this setting would provide a useful complement to the current per-stage and per-long-horizon-task evaluations.

W2 & Q3. Long-horizon failure diagnosis

We thank the reviewer for this observation. We agree that the gap between isolated stage performance and full-task success deserves more detailed analysis. In particular, reporting stage-reaching rates and conditional success after each handoff would help distinguish failures caused by weak component skills, accumulated errors, and transition-state mismatch. We will incorporate these diagnostics in future evaluations to provide a clearer understanding of long-horizon failure modes.

Beyond these controlled studies, we agree that relating method-level differences and scene-specific performance drops to task properties and contact statistics would provide a more informative interpretation of the baseline results. We plan to include such analyses in a future evaluation.

W1. Benchmark variation and generalization

We appreciate the reviewer highlighting this limitation and agree that the current generalization evaluation is relatively narrow. The benchmark includes task-specific randomization of object positions and orientations, as well as robot initial configurations, but these variations are limited in range and are not evaluated as dedicated generalization axes. The reported generalization experiments primarily focus on visual scene changes.

A more comprehensive evaluation should systematically vary object instances and appearance, geometry, layouts, and physical properties. We believe expanding these dimensions would substantially strengthen the benchmark’s ability to evaluate robustness and generalization beyond changes in scene appearance. We also recognize that broader task coverage and quantitative evaluation across additional embodiments would strengthen the scope of the benchmark.

My concerns have not been addressed by this rebuttal. The authors acknowledge issues that should be addressed in a future version of this paper. However, they have not run any additional experiments or provided any new analysis of the current results to alleviate these concerns. I will not be changing my score.

Reviewer_DroyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DexArena is a SAPIEN-based manipulation benchmark containing 10 multi-stage tasks and 30 corresponding stage-level tasks. Its main design choice is to share physical settings and success predicates between isolated stages and their parent tasks, allowing evaluation of individual skills and their sequential composition. The infrastructure supports four end-effectors, state restoration, offline re-rendering, and action replay. The authors release a dataset of 3,000 replay-successful teleoperated stage demonstrations and evaluate PPO, RLPD, ACT, Diffusion Policy, and GR00T on stage execution, three held-out visual scenes, and full-task execution. Stage-level performance is nontrivial, but almost all full-task evaluations fail.

Critical strengths

1,A useful design for studying stage execution and composition in a shared environment. Matching stage definitions, physics, and completion predicates to the parent tasks is more informative than comparing unrelated short- and long-horizon suites. Persistent physical success checks and explicit stage boundaries provide a concrete basis for reproducible evaluation (Sections 3.1 and 4.4).

2,Substantial attention to dataset and simulator reproducibility. The separation of frame restoration from open-loop action replay is particularly valuable: the authors explicitly acknowledge that contact solver state cannot be fully restored and retain demonstrations that succeed under replay. The dataset statistics and replay analysis make the artifact more credible and useful than a collection of task descriptions alone (Section 3.2; Appendices A.5 and A.7).

Critical weaknesses

1,The central long-horizon evaluation has an almost complete floor effect and insufficient diagnosis. Table 16 contains zero success in 48 of 50 task–method entries; ACT achieves only 8% and 4% on the remaining two. Several constituent skills already have near-zero isolated success in Table 2, so full-task failure alone does not identify a composition-specific limitation. The paper does not report stage reach rates or conditional success after actual policy handoffs, nor establish full-sequence feasibility with successful end-to-end expert rollouts. Consequently, it remains unclear how much of the result reflects weak component skills, incompatible transition-state distributions, or genuinely difficult skill composition. This limits the benchmark's demonstrated ability to distinguish long-horizon methods.

2,The evaluated coverage is narrower than the bimanual and heterogeneous-embodiment positioning suggests. Table 5 lists only OpenPackage as a bimanual stage, and Appendix A.3 states that all released demonstrations use the Sharpa Hand. The reported experiments do not compare the four supported end-effectors. Supporting multiple robot interfaces is useful engineering, but it does not establish that the task suite meaningfully evaluates coordination or embodiment variation. These are prominent motivations for the benchmark, so either direct validation or a narrower claim is needed.

Questions for authors

1,Clarification/evidence affecting the long-horizon assessment: Can an expert controller or teleoperator complete every full task without resets or robot repositioning between stages? Please provide end-to-end feasibility results and clarify how handoff states compare with the initial-state distributions of isolated stage demonstrations.

2,Diagnostic evidence affecting the assessment: Can you report the stage reached and conditional stage success in Table 16, especially for sequences whose isolated stages are relatively strong? This would help separate component-skill failures from transition failures.

3,Clarification of scope: Which tasks require simultaneous coordination of both arms, and which have been validated with each supported end-effector? Are demonstrations or evaluation results available beyond Sharpa?

1 comment

We sincerely thank the reviewer for the careful reading and constructive feedback. Below, we provide further clarification on the long-horizon evaluation and handoff states, stage-level diagnostics, and the scope of the current bimanual and cross-embodiment evaluation.

W1 & Q1. Long-horizon feasibility and handoff states

We agree that successful uninterrupted expert trajectories and explicit comparison between handoff states and stage-level initialization distributions would provide stronger validation of the long-horizon evaluation. However, it is beyond our current scope. Please see details in Meta Reply M1.

W1 & Q2. Stage-level diagnostics

We agree that stage-reaching rates and conditional stage success would provide a much clearer diagnosis of long-horizon failures. In particular, these statistics would help distinguish weak component skills from accumulated errors, transition-state mismatch, and failures arising from skill composition. We discuss these diagnostics in detail in Meta Reply M2.

W2 & Q3. Scope of bimanual and cross-embodiment evaluation

We appreciate the reviewer highlighting the distinction between environment/interface support and quantitatively evaluated capabilities. We clarify the current Sharpa-based evaluation, the extent of simultaneous bimanual coordination in the task suite, and future directions for broader cross-embodiment and bimanual evaluation in Meta Reply M3.

Reviewer_Sig1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DexArena is a SAPIEN-based benchmark for dexterous manipulation comprising 10 multi-stage tasks and 30 constituent stage-level tasks. It provides explicit physical success predicates, shared task parameterization, state restoration and re-rendering, and interfaces for a gripper and three dexterous hands. The dataset contains 3,000 replay-successful teleoperated stage-level demonstrations collected with the Sharpa Hand. The benchmark evaluates PPO with demonstration guidance, RLPD, ACT, Diffusion Policy, and GR00T on isolated stages, three held-out visual scenes, and sequential long-horizon execution. GR00T and ACT achieve average stage-level success rates of 61.4% and 58.0%, respectively. Under visual scene changes, their averages are 41.7% and 45.4%. Long-horizon performance is almost uniformly zero: ACT succeeds on furniture assembly (8%) and package opening (4%), while all other method-task entries are zero. The results establish challenging tasks, but leave the causes of the composition failures insufficiently resolved.

Critical strengths
  1. A useful framework for connecting individual skills to full-task execution. The 10 tasks are decomposed into 30 stages with shared physical settings and explicit success predicates, including persistence requirements for many stages (Section 3.1; Table 5, pp. 13–14). This can support more interpretable evaluation than a single end-to-end score, because researchers can examine both skill execution and transitions within a common task definition. The contact-rich tasks span manipulation behaviors beyond simple grasp-and-place.
  2. Substantial attention to demonstration quality and reproducibility. The paper separates frame restoration from open-loop action replay and reports the limitations of the latter. It releases a selected set of 3,000 replay-successful demonstrations from 4,112 accepted episodes, with 3,222 replay successes overall (Sections 3.2 and A.7.3; Table 15). The per-task replay analysis, physical parameter tables, and camera and training details are valuable infrastructure. Reporting replay failures and near-zero long-horizon results also makes important limitations visible.
Critical weaknesses
  1. The central long-horizon evaluation does not yet explain or sufficiently validate composition difficulty. Table 16 (p. 28) contains zero success in 48 of 50 method-task entries, despite oracle switching based on the simulator’s stage-success predicate. There are no reported stage-reach rates, conditional transition success rates, or successful full-task teleoperation controls. Sharing layouts and success predicates does not establish that the terminal states reached by one policy match the next policy’s training initial-state distribution. The current results therefore do not distinguish difficult skill composition from incompatible handoffs, weak component skills, or an unsuitable full-task horizon. Low success is a legitimate benchmark outcome; the concern is that the flagship metric has little resolution and insufficient diagnosis. Reporting intermediate progress and validating complete trajectories on representative tasks would materially strengthen the central claim (Section 4.4; Appendix A.9).
  2. The evaluation protocol needs reconciliation before the baseline comparisons can serve as a dependable reference. Table 10 (p. 21) lists 500-step episodes for PPO/RLPD, whereas Table 5 specifies task-dependent horizons of 300–1,400 steps and Section 4 states that horizons are matched. It is unclear whether 500 applies only during training. Table 10 reports the best checkpoint for RL, while GR00T uses the final checkpoint; the held-out selection procedure for RL is not specified. ACT/DP use evaluation seeds 50000–50049, while GR00T uses 10001–10050. Different seeds are not intrinsically invalid, but they do not provide paired initial states. All methods use one reported training seed, and Table 3’s deviations quantify variation across scenes, not training uncertainty. These details limit interpretation of close rankings and failure comparisons. A reconciled protocol and targeted repeated runs would address this concern (Tables 10–12; Appendix A.6).
Questions for authors
  1. Long-horizon validity and diagnosis. Can you report the full-task horizons, the fraction of episodes reaching each stage, and the conditional success of each transition, together with successful uninterrupted human or reference-controller trajectories on representative tasks? How do next-stage reset states compare with the states actually reached at handoff? These checks would distinguish a composition bottleneck from an evaluation or initialization mismatch.
  2. Protocol clarification. Does the 500-step limit in Table 10 apply only to training, with task-specific limits used for the reported RL evaluations? Which seeds and split were used to choose the RL “best checkpoint,” and were final reported episodes excluded from selection? Would using a common evaluation seed list and a consistent checkpoint-selection rule change the main comparisons?
  3. An informative performance reversal. ACT obtains 0% on isolated FastenScrew in Table 2, 85.3 ± 11.5% across held-out scenes in Table 3, and 8% on full furniture assembly in Table 16. This is not logically contradictory because scene and state distributions differ, but it is unusually informative. Can you confirm that the checkpoint, success test, action mapping, and horizon are consistent, and explain the relevant differences in visual inputs and stage initialization?
  4. Scope of the supported capabilities. Appendix A.3 states that all released data use the Sharpa Hand, and Table 5 identifies only OpenPackage as a bimanual stage. What quantitative task-feasibility or policy results are available for Wuji Hand, XHand, and the gripper? Which tasks test simultaneous bimanual coordination rather than alternating single-arm stages? This would clarify the scope of the cross-embodiment and bimanual claims, and the contribution relative to the related benchmarks in Table 1.
1 comment

We sincerely thank the reviewer for the careful reading and for the concrete diagnostic questions. Below we summarize what the submission does and does not support.

W1 & Q1. Long-horizon validity and diagnosis

We agree that the current long-horizon evaluation would benefit from stronger feasibility validation, explicit analysis of handoff states, and more detailed intermediate diagnostics. We provide a detailed discussion of the stage-level demonstration design and handoff-state issue in Meta Reply M1, and of stage-reaching rates and conditional transition diagnostics in Meta Reply M2. We will also report the full-task evaluation horizons explicitly in a future revision.

W2 & Q2. Protocol clarification
We thank the reviewer for pointing this out. We agree that the current evaluation protocol contains inconsistencies across baselines. We will correct these protocol inconsistencies in the final revised version by adopting aligned evaluation seeds, episode horizons, and checkpoint-selection criteria.

Q3. An informative performance reversal

We thank the reviewer for highlighting this discrepancy. We agree that the FastenScrew results require further verification, and we will re-examine the corresponding evaluation configurations and experiment records to ensure consistency before drawing conclusions from this comparison.

Q4. Scope of the supported capabilities

We appreciate the reviewer’s question regarding the distinction between supported embodiment interfaces and quantitatively evaluated capabilities, as well as between dual-arm task support and simultaneous bimanual coordination. We clarify the current scope of these capabilities in Meta Reply M3.

Comments on the paper

1 comment

We sincerely thank the reviewers for the careful reading and constructive feedback. Several reviewers raised related questions regarding long-horizon demonstrations and handoff states, long-horizon failure diagnosis, and the scope of the current bimanual and cross-embodiment evaluation. We provide the following clarifications that are relevant across these reviews.

M1. Long-horizon demonstrations and handoff states.

We would first like to clarify the role of compositionality in DexArena. Long-horizon skill composition is an important evaluation dimension enabled by DexArena, but it is not the central contribution of the benchmark. DexArena jointly contributes a unified task and evaluation infrastructure, stage-level and long-horizon task definitions, physically grounded success criteria, teleoperated demonstrations, and standardized interfaces for policy evaluation. Consistent with this positioning, the main paper presents stage-level performance and scene generalization in detail, while long-horizon performance is discussed only as an additional evaluation axis, with the results reported at the end of the Appendix A.9.

The current release specifically focuses on stage-level demonstrations for two reasons. First, as a common practice in robot manipulation, collecting and validating complete long-horizon teleoperation trajectories is substantially more demanding in practice than collecting individual stage demonstrations. Second, we intentionally leave the composition of independently learned skills as a future challenge for long-horizon chaining policy evaluation rather than providing complete expert trajectories for every task.

We would like to also explain the reason behind the phenomenon 'close to zero success rate' for long-horizon tasks (Appendix A.9). Each stage demonstration is collected independently from its designed initial state. As a result, the terminal state of one stage demonstration does not necessarily match the initial-state distribution of the following stage, and the transition between the two stages is not represented in the current dataset. The released stage-level demonstrations are therefore not intended to be directly concatenated into full-task trajectories.

In a future version, we plan to complement the current release with complete, transition-consistent long-horizon demonstrations, which will also provide stronger validation of full-task feasibility and support more detailed analysis of policy handoffs.

M2. Long-horizon diagnostics

We agree that the current full-task success metric alone provides limited diagnostic resolution. In particular, the current evaluation does not sufficiently distinguish failures caused by weak component skills from those caused by accumulated errors, inter-stage handoff mismatch, or skill composition.

Reporting stage-reaching rates and conditional stage success after each policy handoff would provide a clearer picture of where long-horizon failures occur. Comparing the states reached at actual handoffs with the designed initialization distributions of the corresponding stage-level tasks would further help identify whether failures arise from transition-state mismatch or from the individual skills themselves.

We agree that these intermediate diagnostics would make the long-horizon results substantially more informative and interpretable, and we plan to incorporate them in future evaluations.

M3. Bimanual and cross-embodiment evaluation

All released demonstrations and quantitative policy results in the current submission use the Sharpa Hand. DexArena provides environments and interfaces for the Wuji Hand, XHand, and gripper, but the current submission does not report quantitative feasibility or policy results for these embodiments. Teleoperation data with multiple hands will be released in the future version.

Regarding the bimanual manipulation, not all of the dual-arm long-horizon task requires simultaneous use of both hands. Most of the tasks primarily compose stages controlled by one arm at a time within a dual-arm environment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.