Dexterous manipulation evidence ledger

Decisions: prioritize native ManiSkill integration (#47) and independent in-hand rotation scoring (#48). Reuse evaluation and input-audit protocols from other methods; defer unqualified runtimes, custom engine forks and hardware-dependent training. Two optional, budgeted follow-ups cover action replay (#52) and synthetic tactile history (#53). Neither precedes core contact/cloth repair.

On 2026-10-05, refreshed 38 historical paper abstract pages and19 repository entrypoints. This is not38 full method reviews or19 reproductions. Unknown source, drive, observation and asset-license fields remain explicit in the paper ledger and repository ledger.

Paper revisions and review depth

The table lists paper identities. Selected method/source audits and abstract-only defer decisions are distinguished below. Unresolved implementation and asset fields block adoption, not the recording of a research decision.

论文 / Paper	版本 / Revision	更新日期 / Revision date
ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI	v2	2025/05/30
DexGraspNet 2.0: Learning Generative Dexterous Grasping in Large-scale Synthetic Cluttered Scenes	v1	2024/10/30
DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning	v2	2025/03/06
TacEx: GelSight Tactile Simulation in Isaac Sim – Combining Soft-Body and Visuotactile Simulators	v1	2024/11/07
Sampling-Based Model Predictive Control for Dexterous Manipulation on a Biomimetic Tendon-Driven Hand	v3	2025/08/04
BODex: Scalable and Efficient Robotic Dexterous Grasp Synthesis Using Bilevel Optimization	v3	2025/09/03
MuJoCo Playground	v1	2025/02/12
Sim-to-Real Reinforcement Learning for Vision-Based Dexterous Manipulation on Humanoids	v2	2025/09/01
ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning	v1	2025/03/27
Taccel: Scaling Up Vision-based Tactile Robotics via High-performance GPU Simulation	v2	2025/09/12
Dexonomy: Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy	v2	2025/09/03
A Survey of Robotic Navigation and Manipulation with Physics Simulators in the Era of Embodied AI	v2	2026/06/09
DexGarmentLab: Dexterous Garment Manipulation Environment with Generalizable Policy	v3	2025/10/12
TeleOpBench: A Simulator-Centric Benchmark for Dual-Arm Dexterous Teleoperation	v2	2025/09/15
DexMachina: Functional Retargeting for Bimanual Dexterous Manipulation	v1	2025/05/30
Dex1B: Learning with 1B Demonstrations for Dexterous Manipulation	v1	2025/06/20
HumanoidGen: Data Generation for Bimanual Dexterous Manipulation via LLM Reasoning	v2	2025/11/16
In-Hand Manipulation of Articulated Tools with Dexterous Robot Hands with Sim-to-Real Transfer	v3	2026/03/05
SPIDER: Scalable Physics-Informed Dexterous Retargeting	v3	2026/09/26
Learning Dexterous Manipulation Skills from Imperfect Simulations	v2	2026/02/25
Dexterous Manipulation Policies from RGB Human Videos via 3D Hand-Object Trajectory Reconstruction	v2	2026/02/11
FlowHOI: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot Manipulation	v1	2026/02/13
Dex4D: Task-Agnostic Point Track Policy for Sim-to-Real Dexterous Manipulation	v1	2026/02/17
Grasp to Act: Dexterous Grasping for Tool Use in Dynamic Settings	v1	2026/02/24
Tacmap: Bridging the Tactile Sim-to-Real Gap via Geometry-Consistent Penetration Depth Map	v2	2026/05/12
HydroShear: Hydroelastic Shear Simulation for Tactile Sim-to-Real Reinforcement Learning	v1	2026/02/28
Structural Action Transformer for 3D Dexterous Manipulation	v1	2026/03/04
PTLD: Sim-to-real Privileged Tactile Latent Distillation for Dexterous Manipulation	v3	2026/06/18
UltraDexGrasp: Learning Universal Dexterous Grasping for Bimanual Robots with Synthetic Data	v1	2026/03/05
Grounding Sim-to-Real Generalization in Robotic Manipulation: An Empirical Study with Vision-Language-Action Models	v2	2026/06/29
Blind Dexterous Grasping via Real2Sim2Real Tactile Policy Learning	v2	2026/06/11
Labimus: A Simulation and Benchmark for Humanoid Dexterous Manipulation in Chemical Laboratory	v2	2026/07/01
Learning Dexterous Manipulation Using Contact Wrench Guidance From Human Demonstration	v2	2026/08/14
TouchWorld: A Predictive and Reactive Tactile Foundation Model for Dexterous Manipulation	v2	2026/07/09
LabDex: A Hierarchical Benchmark for Dexterous Manipulation in Laboratories	v1	2026/08/19
Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds	v1	2026/09/01
Morphology and actuation as inductive biases in robotic hand manipulation	v1	2026/09/04
Dex-X: Learning Visual-Tactile Dexterous Manipulation From Human Videos with Simulated Interaction	v3	2026/10/01

Repository and license evidence

A pinned SHA identifies a source entrypoint, not a paper-matched runnable release. License metadata does not grant asset redistribution rights; absence of a root license does not exclude terms elsewhere.

仓库 / Repository	固定源码 / Source	审查 / Review
mani-skill/ManiSkill	62ff3a5896	README / metadata
JYChen18/BODex	06b9a3c908	README / metadata
JYChen18/DexGraspBench	d9ea6cf282	README / metadata
NVlabs/dexmimicgen	940e8a1b3a	README / metadata
MandiZhao/dexmachina	adae5bf620	README / metadata
Taccel-Simulator/Taccel	cb23bc251b	README / metadata
MMintLab/hydroshear	f815b82fdf	README / metadata
DH-Ng/TacEx	adceed41af	README / metadata
wayrise/DexGarmentLab	e4e298e696	README / metadata
x-robotics-lab/dexscrew	3bde4e3a4d	README / metadata
Genesis-Embodied-AI/genesis-world	02908fe647	README / metadata
InternRobotics/UltraDexGrasp	e0f9784a9b	README / metadata
cyjdlhy/TeleOpBench	4cc410d149	README / metadata
dexx-code/dexx-code	88728c83c9	README / metadata
facebookresearch/project_superdex	1d7150946f	README / metadata
google-deepmind/mujoco	07dfe91655	README / metadata
google-deepmind/mujoco_playground	ef4fefc130	README / metadata
isaac-sim/IsaacLab	a710e671a2	README / metadata
newton-physics/newton	009158e62b	README / metadata

Remaining work

#46 remains open until this research delivery passes review and release. Runtime adoption still requires each record’s unresolved solver, source and asset checks. #47/#48 remain unimplemented.

ManiSkill source audit: upstream success is not physical acceptance

Pinned source: 62ff3a5896b4d5b4cf0ac4c8d79afe600c9404a3. Static inspection only; no new task is running yet. ManiSkill1 v5 emphasizes object generalization; ManiSkill2 v1 extends task/controller diversity. Neither makes current MS3 assets or solvers paper-matched by default.

• PickCube: goal proximity plus robot static defines success. Hold/release and physical support need separate scoring. Non-state extra observations still include grasp status and goal position; RGB selection alone does not prove vision-only input.

• PushCube: planar proximity and an upper height bound are insufficient to reject below-table penetration. Add that explicit negative.

• Rotation: registered Level0–3 use AllegroHandRightTouch. Angle accumulation is unsigned. A mathematical128-step oscillation crosses4π with zero net rotation; this is a formula counterexample, not an executed physics exploit. Score signed palm-relative progress, drops and contact transitions independently. Report its clipped PD effort estimate separately from measured torque.

• Hand controller: PD position/delta targets differ from actual joint state. Touch observations expose contact impulse magnitudes; they are not calibrated tactile images. Audit all observation keys.

Source audit · Algebraic negative. #47/#48 must retain upstream success alongside independent scoring, and test reset isolation. Installed solver identity and asset terms remain qualification requirements.

Synthesis, retargeting and data: reuse decisions

The method ledger records the candidates, source locators, unknowns and proposed budgets. None has been reproduced here.

Method	Verified boundary	Reuse and falsifiable probe
Dexonomy	Historical MuJoCo3.3.3; six-axis loads differ from final-pose checks. Default5cm/15deg is not DexLab acceptance. The hard-mode comment conflicts with execution: noslip2/impratio10 remain; friction decreases.	Reuse grasp taxonomy/loading protocol; compare3 grasp types with fixed object/hand/loads and record directional failures plus effective parameters.
Dex1B	Scale does not establish executable trajectories or contact coverage; implementation/solver/asset audit pending.	Defer bulk data; count generated, executable and independently valid samples, retaining rejected cases.
SPIDER	Pinned README distinguishes saved-state replay from control-replay equivalence; code/data terms separate.	Reuse replay audit: compare free dynamics under fixed actions against pose playback; prohibit object state injection during scoring.
CHORD	AppendixB includes object/contact observations and internal object assistance; whole-body observations differ. Reset warm-up matters even when assistance has zero policy-action dimensions.	Reuse wrench/assistance logging; record every phase and exclude undisclosed assistance from independent acceptance.

Protocol research only: no training or bulk assets. Proposed probes are capped at10 cases, one qualified worker and30min; stop for missing provenance or failed physical scoring. Add separate executable Issues/dependencies before running optional expansions.

Tactile models are not interchangeable physics backends

Work / inspected revision	Reuse boundary	Minimal disproof test
HydroShear v1	Path-dependent shear observation, flat elastomer and compliant-contact assumptions; teacher/student inputs differ.	Same pose through different paths and detach/recontact: check history/reset independently from rigid dynamics.
Tacmap v2	Geometric tactile map; sensing geometry is distinct from physical collision geometry.	Vary sensing resolution only; compare map error and net force separately.
PTLD v3	Requires privileged real-world data collection before tactile deployment; not zero-data transfer.	Audit teacher/critic/collector/actor input keys; remove prohibited pose information from deployment replay.
TacEx v1	Gel physics, optical rendering and marker synthesis are separate modules; historical slipping observations do not indict current PhysX.	Compare force/deformation and optical signals under the same imposed trajectory.
Taccel v2	ABD/FEM IPC and kinematic constraints require separate contact and actuation qualification.	Fixed-mesh indentation/shear/unload; record residuals, force, penetration and stage costs.

A restricted independently implemented geometric occupancy observer now has three bounded experimental batches. This does not adopt or qualify upstream tactile runtimes, shear memory or hardware precision. Upstream input/solver/licensing boundaries remain in the method ledger.

Grasp evaluation and bimanual generation

Method	Evidence boundary	Decision and falsifiable probe
BODex v3	Paper gravity-hold limits are5cm/15deg above3s; geometric penetration and synthesis speed are distinct metrics.	Reuse loading protocol, retain DexLab thresholds; compare identical grasps under frozen physical parameters.
DexGraspBench source	README changes mass30g→100g and kp1→5 relative to BODex.	Record both configurations; isolate mass and drive effects before comparing methods.
DexMimicGen v2	Transformed segments execute open-loop; only successful complete episodes enter the generated dataset.	Preserve attempted, executed and retained counts; score failed trajectories too.
DexMachina source and paper	Installation requires custom Genesis/rl-games forks; training uses decaying virtual object assistance.	Defer direct runtime adoption; reuse assistance logging, with zero auxiliary wrench during independent evaluation.

These are research decisions, not newly supported tasks. Official stable-engine qualification, executable-source review and asset permissions remain prerequisites. Optional probes retain the10-case/30-minute cap; no training has started.

Cloth and tool use: what successful videos do not establish

DexGarmentLab v3 uses particle PBD and adhesion/friction tuning. Its lack of attachment blocks does not establish friction-only grasping; AppendixI still discusses particle gaps, penetration and jitter. Reuse the task taxonomy, but test adhesion separately: fixed mesh/drive, independently varied adhesion/friction, open-hand release and residual attraction. The pinned README targets IsaacSim4.5.0, not our latest-stable qualification.

DexScrew v2 separates a privileged simulation oracle, proprioceptive student, real teleoperation and tactile behavior cloning. This is the paper linked by the repository, not the separate articulated-tools paper. Reuse stage-specific input audits and separate signed rotation, axial progress and full completion. Defer hardware reproduction until real data is available; progress ratio must not be relabeled task success.

Both remain reference-only, with asset and runtime qualification pending. No cloth or tool training has been run for this audit.

Embodiment and actuation

Tendon-hand MPC v3 distinguishes joint targets, tendon commands and estimated joint state. Its VLM tunes objective weights from video while MPC uses measured object state; this is not direct RGB-only control. Reuse that interface audit and test estimation delay and actuator limits separately, recording deadline misses as well as retention.

The morphology study v1 separates kinematics, actuation and their composition. It reports historical MuJoCo3.10.0; its ACBH model and experiment code are available on request, so a full public reproduction is not established. Reuse finite-difference checks of actuator-to-joint and joint-to-tip maps on already licensed models. Structural conditioning cannot substitute for dynamic task acceptance.

Both remain protocol references. Exact source, asset permissions and independent runtime evidence are unresolved; no additional hand model or training is introduced.

Precision, composition and point-based policies

• Labimus v2: reuse separate completion, precision and stage diagnostics. Its dry-powder approximation uses rigid micro-elements; this does not qualify our granular physics. Test a completed but out-of-tolerance negative in an existing task.

• LabDex v1: reuse atomic/compositional boundaries and action/state provenance. Shared real/sim task labels do not prove matched dynamics; exact solver remains unverified. Test a two-stage episode whose second stage fails.

• Dex4D v1: teacher/full geometry and student/partial paired-point inputs differ. Reuse visibility auditing, not a raw-RGB claim. Test target loss without ground-truth replacement.

All three remain reference-only; proposed probes use existing assets with the same10-case/30-minute cap. Full laboratory, granular and policy-training adoption is deferred.

Remaining candidates: screened, not reproduced

These entries were screened from pinned abstracts only. They remain references or deferred implementations; the probes are our proposals, not reported experimental results. Detailed unknowns are retained in the method ledger.

Candidate / 候选	Decision / 决策
MuJoCo Playground	reference runtime design; defer another task integration until ManiSkill qualification
Physics simulator survey	reference taxonomy only; not an engine qualification source
TeleOpBench	reuse operator/interface timing protocol; defer complete benchmark
HumanoidGen	reuse explicit annotation provenance; defer generation stack
Articulated-tool refinement	reuse articulation sensitivity protocol; defer hardware-dependent learning
VIDEOMANIP	reuse reconstruction-versus-execution audit; defer training
FlowHOI	reference HOI representation; defer generative pipeline
Grasp-to-Act	reuse dynamic-load retention protocol; defer learned controller
Structural Action Transformer	reference joint-role representation; defer policy training
UltraDexGrasp	reuse grasp-strategy grouping; defer bulk dataset
Sim-to-real generalization study	reuse factorial calibration design; defer hardware ranking
Blind tactile grasping	reuse contact-event calibration protocol; defer absent hardware data
TouchWorld	reference separate planning/feedback clocks; defer foundation-model stack
Temporal robustness study	reuse speed-conditioned paired evaluation; defer training
Dex-X	reference tactile-supervision provenance; defer executable adoption

Three selected visual/data methods

Source	Reuse decision and boundary
DexGraspNet2.0, §3.2–3.3	Reuse object/clutter splits and proposal accounting. Single-view depth produces a grasp pose, followed by lift evaluation; no continuous RGB-control claim. Compare collision rejection, lift and sustained hold separately.
ManipTrans, §3–4	Reuse effective gravity/friction/threshold logging. Training relaxes physics and termination; freeze nominal evaluation before scoring. Defer training.
Humanoid vision RL, perception setup	Reuse fixed calibrated-camera and tracked-feature provenance. Head depth and third-view segmented3D features are not raw RGB-only inputs. Test occlusion without replacing lost features with simulator truth.

These targeted paper audits do not establish source or asset clearance, installed solver versions or runtime reproduction.