A hand picks up a screwdriver. That first grasp can already decide whether the next step will work: the handle may sit too deep in the palm, the fingers may block a useful contact, or the grip may slip when torque arrives. Dexterity is a sequence of these small decisions. A successful grasp is only the beginning.

CoRL 2026 offers several ways to tackle that sequence. We use Sharpa Wave as the entry point because Play2Perfect, OpenDexGrasp, DEX-X, and WM-Craftnet all use this 22-active-degree-of-freedom hand. This is a selected technical reading, not a conference-wide hardware census: we found no complete, independently verified count supporting the claim that Sharpa was the most-used hand at the conference. [2] [3] [5] [7] [10]

The core comparison covers five main-conference works, adding Touch2Trace on a Tesollo hand. DexAgent is discussed separately as a dexterous-manipulation workshop contribution. SharpaDex is a related dataset release. These publication categories carry different review scopes; a booth demonstration or a release during conference week does not automatically become a main-conference paper. [1] [13] [16] [18]

A shared hand makes the differences easier to inspect

Sharpa Wave combines human-scale geometry, 22 active joints, fingertip sensing, and a software interface with simulation support. That makes it useful for studying both finger gaiting and arm–hand coordination. These are platform capabilities. Whether an experiment actually closes the loop on touch depends on its observation space and controller. [2]

That distinction changes how the conference should be read. Play2Perfect and OpenDexGrasp demonstrate useful capabilities on tactile-capable hardware without using tactile feedback in their deployed policy inputs. DEX-X and WM-Craftnet explicitly use it. A hardware photograph cannot tell us which source of information produced the result. [3] [6] [8] [11]

Work / venueLearning routeWhat the policy observesWhat is actually evaluated
Play2Perfect / mainPlay RL, then assembly fine-tuningProprioception and estimated object/goal posesTask-specialized assembly skills
OpenDexGrasp / mainLanguage-conditioned flow matchingVision, language, object geometryFunctional grasp generation and execution
DEX-X / mainHuman video, simulated teacher, distilled studentPoint cloud, proprioception, touchGrasping and selected tool-use tasks
WM-Craftnet / mainRecurrent world model plus PPODepth, proprioception, touch, action historyIn-hand rotation and robustness tests
Touch2Trace / mainFrozen tactile encoder plus imitationPressure arrays and proprioceptionA cable-tracing skill on Tesollo

Play2Perfect: spend exploration on contact, after learning to play

Play2Perfect pretrains a goal-conditioned reinforcement-learning policy on varied objects and poses, then fine-tunes it on precise assembly. Its paper reports a 33× sample-efficiency gain over training from scratch in the studied setting. This is downstream learning efficiency under its protocol, not a 33× reduction in the full development bill. [3] [4]

On real hardware, tight insertion succeeds in 6/10 trials at 0.5 mm clearance; successful runs average 11.1 ± 5.1 seconds. The controller runs at 60 Hz, uses object-pose estimates, and transfers without real-world policy fine-tuning after task-specific simulation training. The two beam-assembly stages were tested separately. The paper also reports seven days of play pretraining and one day per downstream policy on an RTX A6000. [3]

The industrial attraction is clear: an adaptable hand could reduce the need for a new fixture or attachment for every part. The unresolved cost is equally concrete. A new assembly still needs task setup and fine-tuning; pose tracking can fail under occlusion, and the controller does not directly observe surrounding fixture geometry. We would evaluate it as a reusable skill-training recipe, with each new production task earning its own acceptance test.

OpenDexGrasp: choose contacts that leave the tool usable

OpenDexGrasp asks how language should change a grasp. Holding a pitcher for pouring and holding it for handover call for different contacts. Its model couples visual-language features with metric geometry, using flow matching to generate functional hand poses. Auxiliary affordance supervision teaches which object regions matter without requiring a separate affordance-to-pose cascade at inference. [5]

The data recipe combines 1.24 million automatically synthesized grasps with a smaller alignment set of 27.42 thousand teleoperated and transferred poses. On a Franka Panda plus Sharpa Wave, the reported real-world average is 72%, versus 59% for the adapted DexGraspNet 2.0 baseline, across seen and unseen categories. The paper states that execution is open-loop and tactile correction is future work. [5] [6]

This fills an earlier gap in the workflow: decide where the fingers should go before trying to regulate contact. A functional grasp can make the next action easier, but the grasp result does not establish that the robot can complete pouring, fastening, or another long task. We would preserve the grasp generator as a proposal module and test the downstream controller independently.

DEX-X: simulation supplies contact that a video cannot record

DEX-X starts from monocular human demonstrations and reconstructs hand–object interaction in simulation. A privileged reinforcement-learning teacher tracks the references with contact information; dataset aggregation distills a deployable student reading point clouds, proprioception, and touch. The released platform uses a Franka FR3 and Sharpa Wave. The transfer is zero-shot to hardware, with no robot-side demonstration collection in the reported recipe. [7] [9]

Its real cube-picking policy succeeds in 28/30 trials; removing touch reduces that to 11/30. Table cleaning reaches 53% over 30 trials. Generalization is uneven: the unseen thin cube reaches 23/30, while the square cube and large duck each reach 8/30. The simulation teacher's 65.9% average across six categories is a separate metric, not a real-world average. [8]

The method makes human video more useful by adding a physical learning stage. Its weakness also sits there: reconstructed geometry, friction, and sensor simulation can be wrong. Our reading is that the contact model becomes part of the data-production system. A team needs to audit this synthetic supervision with the same care it would apply to a human annotator.

WM-Craftnet: keep a memory of contact while the object moves

WM-Craftnet combines noisy wrist depth, tactile contact, proprioception, and action history in a recurrent World Synesthesia Model. An asymmetric actor–critic policy uses that model's state. It is a world-model-conditioned controller: the deployed model estimates task-relevant state, rather than planning by searching imagined video rollouts. [10] [11]

On Sharpa Wave, the experiments cover multi-object rotation, unseen geometry, and perturbation recovery. Hardware success means more than half a turn within a 20-second episode, so it measures a bounded in-hand skill. The paper also shows that masking all tactile input at test time lowers simulated return from 753.3 to 724.9. This modest drop cautions against attributing the full gain to touch alone; recurrence and clean-depth supervision matter too. [11]

This is a useful corrective to the idea that a world model must generate photorealistic futures. For contact control, a compact memory of the hand–object state may be more useful than a beautiful prediction. The release includes training scripts, a z-axis checkpoint, and deployment examples, although checkpoint heads and current training defaults must be matched before reproducing a result. [12]

Touch2Trace: a different hand exposes the timing problem

Touch2Trace uses a Tesollo DG-5F with custom pressure arrays, not Sharpa. A frozen tactile encoder feeds a lightweight temporal imitation policy. Twelve task demonstrations amount to 10.1 minutes, while encoder pretraining uses a separate contact corpus. Counting only those minutes would understate the full data requirement. [13] [14]

On an unseen Ethernet cable in straight routing, it reports 20.1 cm mean tracing distance and 93% success at reaching 10 cm, versus 57% at reaching 20 cm. Each condition has 30 trials; aborted starts shorter than five seconds are excluded. This is transfer across cable and routing conditions within one skill, not general cable assembly. [14]

The most useful engineering result may be latency. Its temporal policy runs at 60 Hz with 3.4 ms model inference. The tested diffusion policy takes 32 ms for ten denoising steps and performs poorly at the matched 60 Hz setting; at 30 Hz it recovers some performance. Architecture, inference budget, and training settings all matter here. This experiment does not establish that diffusion is generally unsuitable for touch. [13] [14]

For a hand vendor or integrator, that is an actionable question: does the sensing-to-actuation path fit the control period after encoding, communication, and inference? A high sensor frame rate alone cannot answer it. Neither can a nominal SDK communication rate. The entire loop must be measured under load.

SharpaDex: data infrastructure deserves its own scrutiny

The related SharpaDex release provides 300 hours, 28,993 episodes, and 59 real teleoperation tasks with synchronized vision, joint data, and fingertip touch. Its dataset card is explicit about storage details: the 65-dimensional schema includes seven torso dimensions that are zero, and v3.0 and v2.1 are exports of the same demonstrations. They must not be added together. [15]

This is useful infrastructure for learning which contacts accompany a successful action. It does not by itself establish policy reliability, nor does a demonstration corpus automatically cover failures and recoveries. The stated dataset license is CC BY-NC 4.0. Industrial teams should establish permitted use before making it a dependency in a commercial training pipeline. [15]

DexAgent and the workshop boundary

The dexterous-manipulation workshop lists DexAgent as a spotlight contribution. It uses an agent and a growing library of skills and verifiers to reconstruct human demonstrations, generate simulation trajectories, and produce policy data. The project reports 63.6% average learned-policy success across eleven real tasks, compared with 18.2% for its strongest baseline. [16] [17]

The same page reports open-loop replay success on 11/11 tasks, with a task counted if at least one of ten trials succeeds. That measures coverage of trajectory conversion, not 100% policy reliability. The workshop direction is interesting because it automates parts of the data workflow; every generated verifier still needs validation before its judgments become rewards or release gates. [17]

What connects the papers—and what cannot be ranked

These works address different failure points. OpenDexGrasp proposes useful contacts. Play2Perfect learns how to refine interaction near an assembly goal. DEX-X recovers missing contact supervision from human video. WM-Craftnet maintains state through contact changes. Touch2Trace tests how quickly and in what form tactile information must reach a controller. Their percentages answer different questions.

Putting them on the same Sharpa platform would help control one source of variation, but would not create a fair leaderboard. Cameras, task resets, object knowledge, policy inputs, and training compute would still differ. Combining their modules is a research proposal; none of these papers establishes that the pieces work together out of the box.

Our deployment assessment is controlled experimental demonstrations, with meaningful real-hardware results in bounded tasks. The reviewed evidence does not establish shift-long uptime, maintenance cost, calibrated damage rates, or a common intervention metric. A ten-trial insertion result and a short rotation episode cannot settle those questions.

For an industrial pilot, start from a failure that costs money

Choose a recurring task: a part jams during insertion, a tool rotates in the palm, or a cable escapes the fingers. Measure the existing process, including fixture setup, operator resets, and damaged parts. Then select the research module that addresses that failure, and test it on held-out objects and realistic disturbances under a fixed adaptation budget.

We would track complete-task success, cycle time, recovery, human intervention, and fingertip maintenance together. Record camera and tactile latency, actual control frequency, training cost, and the data needed for each new part. A hand's industrial value becomes clear when these numbers improve together over repeated runs.

Sharpa is a useful lens on CoRL 2026 because the same hardware appears in several serious attempts to solve different parts of dexterity. The progress is tangible: better functional grasps, reusable manipulation priors, and policies that can respond to contact. We would follow these tools through a bounded pilot, while keeping hardware adoption and deployment readiness as separate claims.

Connect richer data to an executable feedback path

PolyUMI and Eucalyptus ask what contact information can be recorded; the Sharpa studies ask how a policy uses related information on a robot hand. Those sensors and embodiments are not interchangeable. AW-0 adds program structure and visual foresight, which address a different failure class. Combining these directions is a research proposal: demonstrate fewer autonomous contact failures at matched latency before attributing dexterity to a larger agent or richer dataset.

Primary sources

  1. CoRL 2026: conference and workshop dates
  2. Sharpa Wave: platform specifications
  3. Play2Perfect: paper and real-world protocol
  4. Play2Perfect: project and control demonstrations
  5. OpenDexGrasp: CoRL oral, data, and platform
  6. OpenDexGrasp: evaluation and limitations
  7. DEX-X: CoRL acceptance and pipeline
  8. DEX-X: real-world ablations and generalization
  9. DEX-X: released hand–arm implementation
  10. WM-Craftnet: CoRL spotlight and project
  11. WM-Craftnet: predictive-state and tactile ablations
  12. WM-Craftnet: code, checkpoint, and configuration caveats
  13. Touch2Trace: CoRL project and timing
  14. Touch2Trace: data, trial exclusions, and metrics
  15. SharpaDex v1.0: dataset card and license
  16. CoRL dexterous-manipulation workshop: accepted contributions
  17. DexAgent: replay and learned-policy evaluation
  18. Stanford IPRL: Play2Perfect CoRL acceptance