PaperlayerView exact revision on arXiv

Source-linked AI summary

ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning

Jayjun Lee, Jessica Yin, Asif Rana, Nicholas Blauch, Sam Mady, Mohak Bhardwaj, Nima Fazeli, Nathan Ratliff, Karl Van Wyk, Ankur Handa

arXiv:2608.19182v1cs.ROcs.AI

TL;DR

Dexterous robot learning struggles to transfer skills across long-horizon tasks without rediscovering low-level manipulation behaviors. ADEPT pretrains reusable dexterity, stabilizes downstream reinforcement-learning adaptation, and transfers distilled policies zero-shot from simulation to two real arm-hand embodiments.

  • Problem

    High-DoF, contact-rich dexterous systems remain difficult to train, while task-specific reinforcement learning rarely transfers and repeatedly rediscovers foundational manipulation skills.

  • Method

    ADEPT pretrains arm-hand-object dexterity on generic object reposing, then uses structured post-training, distillation, and joint-space geometric fabrics for downstream adaptation.

  • Results

    ADEPT supports long-horizon grasping, reorientation, transport, and insertion across Kuka-Allegro and Flexiv-Sharpa, transferring through vision and visuo-tactile students to real robots.

  • Takeaways & Limitations

    Reusable dexterous pretraining can reduce the manipulation behavior that must be rediscovered for each new task while preserving downstream adaptation.

  • Takeaways & Limitations

    Robust object-centric perception under occlusion remains an open limitation, with failures linked to incorrect asymmetric-peg orientation estimates and grasp instability.

Abstract

from arXiv · show

We introduce Accelerating Dexterity via Pre-Training (ADEPT), a large-scale reinforcement learning (RL) framework for learning sim-to-real transferable dexterity across high degree-of-freedom (DoF) robot embodiments that can solve long-horizon tasks directly from raw visuo-tactile perception. ADEPT pretrains a dexterous policy on a generic object reposing task, then post-trains downstream policies with this pretrained behavior as a prior. ADEPT enables learning new behaviors that are otherwise difficult to discover from scratch on multi-fingered robots and avoids learning the same set of skills over again for every new downstream task. The pretrained policy zero-shots the reposing phase of downstream tasks, but naïve RL fine-tuning rapidly degrades this capability during transfer. We address this with a stable post-training recipe combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates. To safely exploit the full kinematic dexterity, we introduce a joint-space Geometric Fabric that mediates between the RL policy and the robot. We distill post-trained teachers into perceptive students that zero-shot sim-to-real transfer on two embodiments: a 23 DoF Kuka-Allegro with two RGB cameras, and a 29 DoF Flexiv-Sharpa with two RGB cameras and five vision-based tactile sensors, and can solve long-horizon tasks from challenging initial states with dexterity at human-level speed.

1 Introduction

ADEPT addresses the difficulty of learning transferable dexterity in high-DoF, contact-rich robot systems by pretraining on generic object reposing and post-training for downstream tasks. Its stable RL adaptation, full joint-space safety mechanism, and distillation curriculum enable zero-shot sim-to-real dexterous manipulation across multiple embodiments.

  • Motivation: High-DoF arm–hand systems remain difficult to train because large state and action spaces and contact-rich interactions make useful behaviors hard to discover from sparse task-specific rewards.RL policies trained from scratch for one task rarely transfer, motivating a general-purpose dexterous foundation followed by downstream adaptation.
  • ADEPT framework: ADEPT pretrains foundational arm–hand–object dexterity on generic object reposing, then post-trains the policy for contact-rich downstream tasks.The pretrained policy zero-shots downstream reposing segments, while naïve fine-tuning can rapidly degrade that capability during transfer.
  • Safety and dexterity: The full joint-configuration-space geometric fabric exposes high-DoF kinematic dexterity while preventing collisions and joint-limit violations.It contrasts with prior fabric-guided policies that restrict the hand to a low-dimensional PCA grasp subspace.
  • Sim-to-real deployment: A structured post-training recipe and distillation curriculum enable zero-shot sim-to-real for long-horizon dexterous insertion across a vision-based Kuka-Allegro and visuo-tactile Flexiv-Sharpa.The approach uses post-trained teachers and deployable end-to-end perceptive students without demonstrations or pose trackers.
  • Execution speed: 2×–14× execution-time speedup over the FMB parallel-jaw pipeline is achieved, with multi-fingered policies solving each task in 5–10 s versus 20–70 s per trial.The FMB pipeline relies on external fixtures and multi-stage regrasp decomposition.

2 Related Work

Related work spans sim-to-real dexterous manipulation, reinforcement-learning pretraining and transfer, geometric fabrics, and tactile sensing. ADEPT is positioned as combining these directions for high-DoF dexterous policy learning and transfer.

  • Sim-to-Real Dexterous Manipulation: Prior work established large-scale PPO with domain randomization for sim-to-real dexterity and pursued in-hand reorientation.ManipGen composes local simulation-only policies with motion planners, while OmniReset diversifies reset-state distributions to elicit emergent dexterity from PPO.
  • Reinforcement Learning Pretraining and Transfer: Prior RL studies use pretraining, behavior priors, and regularization to mitigate distribution shift or forgetting during online adaptation.ADEPT instead targets high-DoF dexterous policies trained from scratch in simulation and initializes downstream actors through behavior distillation rather than explicit KL or EWC penalties.
  • Geometric Fabrics for Policy Learning: Geometric fabrics provide smooth, second-order action priors with provable stability guarantees and built-in collision and joint-limit avoidance.Earlier dexterous applications used fabrics for sim-to-real grasping but restricted hand control to a 5D PCA subspace of retargeted human motion.
  • Tactile Sensing for Sim-to-Real Dexterous Manipulation: Sim-to-real dexterous policies have largely relied on vision because simulated contact signals transfer less cleanly than images.Recent approaches bridge the gap through shared representations: HydroShear simulates hydroelastic shear for vision-based tactile sensors, while TacMap represents tactile observations.

3 Methodology

ADEPT uses a three-stage pipeline that pretrains dexterity on generic reposing, stabilizes downstream RL post-training, and distills task-specialist behavior into a zero-shot deployable vision-based student. The method preserves pretrained manipulation skills during transfer while enforcing robot constraints through a geometric fabric.

  • Three-stage framework: ADEPT first pretrains a dexterous policy with PPO on generic object reposing, then post-trains it for contact-rich downstream tasks and distills the result into a stereo RGB student.The pipeline comprises pretraining, stable RL adaptation, and teacher-student distillation for zero-shot real-robot deployment.
  • Dexterity pretraining: The generic pretraining task spans reaching, grasping, lifting, in-hand reorientation, transport, and reposing across 16 randomized primitive shapes and scales.The task is designed to instill broad manipulation capabilities, with ADR providing an online curriculum.
  • Transfer challenge: Standard RL fine-tuning rapidly destroys pretrained reposing behavior, reaching 0% success after some updates despite strong initial zero-shot performance.The degradation is attributed to reward and observation mismatches, poor value or advantage estimates, and large policy updates.
  • Structured RL adaptation: ADEPT treats downstream learning as layering behaviors onto the pretrained prior, restricting on-policy PPO to local policy-space updates while retaining core manipulation skills.The adapted policy incorporates downstream task-specific observations, including receptacle pose and object–receptacle contact forces, together with the new reward.
  • Student distillation: After post-training, DAgger distills the state-based teacher into a student using proprioception, fabric state, two RGB images, and five per-finger tactile maps on Flexiv-Sharpa.The resulting student is intended for zero-shot deployment on the real robot.
  • Student curriculum: A two-stage student curriculum separately trains visual perception with an 8-keypoint pose auxiliary loss before refining the contact-rich policy, avoiding perception-policy objective conflict.The auxiliary loss addresses peg-orientation perception, a dominant insertion failure mode.
  • Geometric fabric: A geometric fabric mediates between policy and robot to enforce joint-limit and collision constraints, shape exploration with a smooth second-order prior, and use the same low-level controller in simulation and reality.The fabric is formulated as an autonomous second-order system on configuration space.

4 Experiments

ADEPT transfers reposing skills to unseen objects and downstream manipulation, while its stable post-training recipe avoids naïve fine-tuning collapse and enables novel behaviors. Real-world students solve multi-stage tasks with vision or visuo-tactile sensing at human-comparable speed.

  • Teacher generalization: The pretrained teacher generalizes from 16 primitive shapes to FMB pegs and VisDex objects, matching or slightly exceeding in-distribution reposing success on both OOD sets.Evaluation covers Kuka-Allegro and Flexiv embodiments across three object categories.
  • Zero-shot downstream transfer: For FMB peg insertion, the pretrained policy reliably zero-shots reposing, leaving only contact-rich insertion to learn; for dish placement, transfer reaches but does not grasp.The dish geometry is unlike the 16 primitive pretraining objects, although post-training later learns the full task.
  • Training and stability: 11B total environment steps yield an ADEPT downstream teacher after 8B-step reposing pretraining and 3B-step post-training, whereas naïve PPO rapidly collapses to zero success.Naïve PPO fails because the pretrained critic is calibrated to reposing rewards, making insertion-task advantages unreliable.
  • Real-world results: A single Kuka-Allegro reposing checkpoint post-trains to three tasks, achieving 5/10, 3/10, and 6/10 real-world success, while Flexiv-Sharpa reaches 8/10 on square/round insertion.The Kuka tasks are FMB star insertion, FMB square/round insertion, and dish-rack placement; no task-specific pre-training is used.
  • Real-world results: Visuo-tactile sensing reaches 8/10 overall versus 3/10 for vision-only Flexiv-Sharpa, because tactile feedback resolves grasp confidence and prevents drops and regrasping loops.The largest vision-only stage failures occur during lifting and reorienting.
  • Human-level speed: Policies execute continuous one-hand manipulation in 5–10 s per trial, compared with 20–70 s for fixture-dependent human demonstrations using a parallel-jaw gripper.This corresponds to a 2×–14× execution-time difference.

5 Limitations

ADEPT’s real-world behaviors closely match simulation, but perception remains the primary distillation bottleneck. Failures stem mainly from occluded peg-orientation errors and unstable grasps during fast arm motion, with fingertip tactile sensing mitigating contact ambiguity on Flexiv-Sharpa.

  • Perception remains the primary bottleneck for distilling real-world policies, despite behaviors closely matching simulation.The limitation concerns distillation rather than the overall simulation-to-real behavioral match.
  • Incorrect estimates of the asymmetric peg’s orientation under occlusion frequently coincide with failures.
  • Fast arm motion can destabilize grasps because rounded Allegro fingertips contact the peg over a small patch.
  • Fingertip tactile sensing on Flexiv-Sharpa mitigates contact ambiguity.

6 Conclusion and Discussion … Primitive Objects for Pre-Training

ADEPT combines generic reposing pre-training with structured post-training to preserve transferable dexterous behaviors while adapting to contact-rich downstream objectives. The accompanying task specification defines reposing rewards, contact gating, difficulty annealing, and randomized primitive objects.

  • 6 Conclusion and Discussion: ADEPT uses generic reposing pre-training and structured post-training to initialize downstream manipulation while preserving useful pretrained behaviors during adaptation.The framework supports long-horizon behaviors across Kuka-Allegro and Flexiv-Sharpa, including grasping, reorientation, and transport.
  • A Task MDPs: The reposing task combines hand-to-object proximity, object-to-goal alignment, and contact-gated reward components.The supplied reward specification explicitly includes object-to-goal alignment and contact gating terms.
  • A.1 Reposing Task: Reward component weights are wr = 1, wg = 5, and wb = 0.01.These weights correspond to the reposing reward components listed in the task specification.
  • A.1 Reposing Task: The hand-distance term uses αr = 10 and the maximum Euclidean distance between the object and all hand bodies.dh is defined as maxj ∥phand j − p∥2.
  • A.1 Reposing Task: The goal-alignment sharpness parameter changes from αg: 15 →30 (ADR), using mean distance across eight cube bounding-box keypoints.The cube has half-extent h = 0.15m.
  • A.1 Reposing Task: Contact gating activates when the thumb and at least one other finger exert force above τ = 1N.The gate gc is binary, with gc ∈ {0, 1}.
  • Domain Randomization (ADR): Difficulty is annealed over 50 ADR increments when running success exceeds 0.4, including gravity and object-scale randomization.Gravity changes from 0 to −9.81 m/s2 and object scale from 0.5–1.0×; success requires ekp < 0.10 m within 4 s.
  • Primitive Objects for Pre-Training: Pre-training uses 16 primitive objects whose scale and physical properties are randomized.These objects constitute the primitive-object set for the reposing task.

A.2 Downstream Tasks from Scratch and Post-training … B Joint Configuration Space Geometric Fabric Setup

The downstream pipeline uses identical rewards for scratch and bootstrapped policies, then combines reposing pretraining, behavior-cloning warm-start, critic warm-up, conservative PPO, and vision distillation. The resulting policies use staged observation sets and a 60 Hz joint-space geometric fabric across 23-DoF and 29-DoF embodiments.

  • A.2 Downstream Tasks from Scratch and Post-training: Both training-from-scratch and bootstrapped downstream policies use identical reward parameters; only initialization differs.The shared weights are wr = 1, αr = 10, wg = 5, and αg = 15, with finger-curl and lift terms disabled.
  • A.2 Downstream Tasks from Scratch and Post-training: Goal tolerance anneals from 0.05 →0.02 m over 50 ADR increments, advancing when success rate exceeds 0.4.An episode succeeds when epos < ϵgoal.
  • A.3 ADEPT Pre-training and Post-training Algorithm: Post-training reduces the actor learning rate by 100× and tightens the PPO clip by 4×, while retaining the critic learning rate and switching to linear decay.Pretraining uses a KL-adaptive actor schedule; post-training uses linear decay over Max epochs.
  • A.3 ADEPT Pre-training and Post-training Algorithm: ADEPT trains a reposing teacher with PPO, initializes the downstream actor through behavior-cloning distillation, warms up the critic with the actor frozen, and then applies PPO post-training.The final stage unfreezes the actor and updates actor and critic using clipped PPO.
  • A.4 Observations per stage.: The observation pipeline progresses from privileged-state reposing pretraining to receptacle-aware downstream bootstrapping and two-camera RGB vision distillation.The Stage 2 actor adds receptacle pose and object–receptacle contact, while the Stage 3 student replaces point clouds with left and center RGB images.
  • A.4 Observations per stage.: For Flexiv–Sharpa, Stage 3 distills the state-based teacher into a stereo RGB plus five-fingertip TacMap student.The embodiment has 29 DoF, and the pipeline uses the FMB bootstrap teacher before distillation.
  • B Joint Configuration Space Geometric Fabric Setup: The geometric fabric combines a positive-definite system metric, nominal path-generating force, and policy-driven forcing while integrating policy targets at 60 Hz.The fabric state evolves in joint configuration space, and second-order Runge–Kutta integration is also performed at 60 Hz.
  • B Joint Configuration Space Geometric Fabric Setup: The fabric partitions joint-space targets into arm and hand slices for both the 23 DoF KUKA-Allegro and 29 DoF Flexiv-Sharpa systems.The KUKA-Allegro configuration is 7 arm + 16 finger DoF, while Flexiv-Sharpa is 7 arm + 22 finger DoF.

B.1 Pull-Back from Taskmaps to the Root … B.8 ADR-Annealed Fabric Parameters

The Geometric Fabric pulls taskmap forces and metrics into root space, while combining attraction, collision and joint-limit avoidance, damping, speed control, and bounded action integration. Its parameters are incorporated into ADR so regularization gradually relaxes as policy competence improves.

  • B.1 Pull-Back from Taskmaps to the Root: Taskmap contributions are pulled back to root space using Jacobians, with a curvature force preserving the chain-rule-consistent leaf acceleration.Each component contributes a leaf metric and force; the pull-back includes the curvature term M ˙J ˙qf.
  • B.2 Cspace Attractors: Cspace attractors combine HD1 forcing with soft radius damping and an HD2 geometric attractor, producing critical damping near targets and speed-invariant paths.The geometric attractor uses the same direction with quadratic-in-velocity scaling and does not bias steady-state convergence.
  • B.3 Body Sphere Collision Avoidance: The fabric models the robot with Ns = 31 collision spheres and applies separately normalized, budget-weighted forcing and geometric responses through per-sphere Jacobian pull-backs.Per-sphere pull-back prevents aggregate normalization from diluting budget weighting as the number of spheres grows.
  • B.4 Joint Limit Repulsion: Joint-limit repulsion normalizes signed clearance by joint range, uses a smooth retreat gate to suppress chatter, and adds post-hoc quadratic-program acceleration and jerk caps.The barrier retains unbounded 1/(x−δJ)^2 behavior near limits while tapering during retreat.
  • B.5 Cspace Damping and Speed Control: Cspace damping and energy-based speed control regulate motion smoothness, with energy defined over arm and palm taskmaps while hand joints remain outside the energization budget.The speed-control damping engages only above E⋆= 1.0, with bsc = 100, yielding configuration-consistent end-to-end speed.
  • B.6 Action-to-Cspace Mapping: Actions are bounded relative joint deltas scaled by δmax = 0.1 rad, using the fabric’s integrated configuration as the policy target while retaining the full 23-DoF action manifold.The relative formulation avoids large absolute-target jumps while allowing cumulative trajectory-wide dexterity.
  • B.7 Policy–Fabric Integration Loop: Each target is broadcast to the fabric, which performs K = 2 default inner ODE integration steps before issuing state outputs to simulated PD or real-robot admittance controllers.CUDA-graph capture keeps per-step cost negligible at the parallel scale required for large-batch RL.
  • B.8 ADR-Annealed Fabric Parameters: ADR anneals cspace damping, the speed-control energy target, and velocity attenuation, gradually relaxing fabric regularization as policy competence improves.Damping moves from heavily smoothed toward reactive control, while the energy target broadens exposed motion speeds.

C Population-Based Training

The reposing pretraining teacher uses decentralized Population-Based Training to search the joint PPO and ADR hyperparameter space. PBT ranks parallel policies by unshaped success and combines periodic replacement, ADR-preserving transfer, and conservative hyperparameter mutations.

  • Population-Based Training: Decentralized PBT trains the reposing teacher by searching the joint PPO + ADR hyperparameter space.Policies are trained in parallel to improve hyperparameter search efficiency.
  • Ranking and replacement: PBT launches N = 16 policies, begins checks after Tstart = 200M environment frames, and repeats them every TPBT = 200M frames.Comparisons are restricted to checkpoints with approximately equal frame counts before ranking by the target objective.
  • ADR is preserved on replacement: Replacement preserves the receiving policy’s ADR level, combining the donor’s well-shaped weights with an easier curriculum slice.This lets the rewritten policy approach the population frontier without rediscovering basic dexterous behaviors.
  • Mutation: At each PBT check, PPO hyperparameters are independently mutated with probability pmut = 0.25 through log-uniform perturbations.The mutations cover learning, clipping, entropy, critic, bounds-loss, KL, mini-epoch, discount, and GAE settings, with specified clamps for PPO clip ϵ and mini-epochs.
  • Target objective: Policies are ranked by running success rate at their current ADR level, using the simulator’s true objective info rather than shaped reward.This compares actual task progress despite shaped-reward changes caused by PPO hyperparameters and ADR levels.

D Post-Training Recipes

The post-training recipe stabilizes transfer by first distilling the teacher into the actor, then warming up the critic with a frozen actor before conservative PPO updates. Task-specific FMB insertion shaping preserves necessary contact while adapting exploration, goals, and observations for post-training.

  • Stage 2 — Actor warm-start via BC distillation: BC for 40,000 iterations converges the student action distribution to the teacher’s while mixed student-teacher rollouts keep training states deployment-relevant and teacher-anchored.Only the actor is optimized with a Mahalanobis-weighted distance; the critic remains unchanged during BC.
  • Stage 3 — Critic warm-up with frozen actor: For the first 20 post-BC epochs, the actor is frozen while the critic updates against rpost, aligning value estimates with Mpost before policy-gradient updates.The goal is set to the final insertion endpoint at ADR 50, while ADR randomizations continue advancing.
  • Stage 4 — Conservative PPO post-training: Actor learning rate drops to 1×10−5 with linear decay over 200,000 epochs, while PPO clipping tightens to ϵ = 0.05 for bounded policy updates.The actor–critic uses separate trunks, and observation clipping increases from 5.0 to 100.0 for larger post-training contact and pose magnitudes.
  • Task-specific reward shaping for FMB peg insertion: FMB peg insertion sets both peg–board and receptacle contact penalties to 0 because insertion requires sustained contact and the pretrained policy already handles contacts gracefully.External object wrenches are disabled because the peg remains grasped throughout post-training.
  • Task-specific reward shaping for FMB peg insertion: Peg orientation randomization starts at its maximum range ([1.0, 1.0]), while object-to-goal reward sharpness is fixed at 15 rather than annealed.These settings are applied from the start of post-training instead of being ramped via ADR.

E ADEPT Training Efficiency … G.3 Total Distillation Objective

ADEPT post-training converges reliably compared with scratch training, whereas direct PPO fine-tuning can collapse success during transfer. Its real-world deployment preserves the simulation interface, and its student distillation objective combines moment-matching behavior cloning with auxiliary peg-pose prediction weighted by wBC = 1 and waux = 20.

  • E ADEPT Training Efficiency: Direct PPO fine-tuning of the pretrained reposing teacher exhibits success-rate collapse during transfer on FMB peg insertion.
  • E ADEPT Training Efficiency: ADEPT post-training reliably converges, while scratch-training seeds often plateau or require very long training under a fixed compute budget.
  • F Real-World Deployment Topology: Real-world ADEPT preserves the exact simulation action interface and training observation graph, replacing only simulated physics and rendering with hardware controllers and sensors.
  • F Real-World Deployment Topology: Two Intel RealSense cameras provide 320 × 240 images over ZMQ, while Jetson-hosted control and fabric processes connect to a separate workstation running the student policy.
  • F Real-World Deployment Topology: The deployment uses a 1 kHz C++ admittance controller, a 1 kHz KUKA loop, a 333 Hz Allegro loop, and 60 Hz Python-side processes.
  • G Distillation Loss Details: The distillation objective uses a stochastic Gaussian student with mean µθ(ot) and standard deviation σθ(ot), supervised by a teacher with mean µT(oTt) and fixed standard deviation σT.
  • G.1 Behavior-Cloning Loss: Behavior cloning matches both action-distribution moments using a Mahalanobis-weighted L2 distance, with teacher uncertainty providing per-dimension precision weighting for nq = 23.
  • G.2 Auxiliary 8-Keypoint Pose Loss: The auxiliary head predicts eight peg bounding-box corners in world coordinates using per-corner 3D L2 residuals against simulator-derived ground-truth positions.

H Visuo-Tactile Student Architecture · I Distillation Domain Randomization

The student architecture combines two-camera vision, per-finger tactile sensing, proprioception, fabric state, and recurrent policy processing, with the KUKA-Allegro variant removing the tactile branch. Teacher–student distillation uses stage-matched aggressive domain randomization to support zero-shot sim-to-real transfer.

  • H Visuo-Tactile Student Architecture: The Flexiv-Sharpa student consumes two-camera RGB observations, noisy proprioception, fabric state, and per-finger TacMap depth maps.The KUKA-Allegro vision-only student is the same architecture with its tactile branch removed.
  • H Visuo-Tactile Student Architecture: A shared ResNet and cross-attention fuser encode the camera pair into a 256-dimensional vision latent.The backbone is initialized from the Stage-1 pretrained student and unfrozen on the DextrAH-RGB schedule.
  • H Visuo-Tactile Student Architecture: Each tactile depth map is processed by a shared CNN, while fingertip-position FiLM modulation softly anchors the per-finger features.The modulation uses α = 0.1, and the five resulting features form a tactile latent with dtac = 160.
  • H Visuo-Tactile Student Architecture: Vision, tactile, proprioception, and fabric latents feed [512, 512] MLP layers and a layer-normalized 1024-unit LSTM before the action heads.The final head emits a mean action with fixed log-standard-deviation matching the teacher’s parameterization.
  • H Visuo-Tactile Student Architecture: Pure teacher behavior cloning trains the student with an auxiliary 8-keypoint peg-pose target on Flexiv-Sharpa FMB.The peg-tight-keypoints auxiliary term uses coefficient waux = 20; other exposed auxiliary heads are disabled for this experiment.
  • I Distillation Domain Randomization: Both teacher–student distillation stages apply aggressive randomization across physics, scene layout, sensor noise, and visual rendering.Each stage uses the randomization schedule of its corresponding RL stage, matching the teacher’s environmental distribution.
  • I Distillation Domain Randomization: Visual randomization includes a sample 16x16 grid of camera variations in the FMB training scene.The example concerns the left camera used for distillation.
  • I Distillation Domain Randomization: Table 8 organizes distillation randomization into per-environment reset ranges, ADR-annealed ranges, per-step noise, and per-frame camera-render randomization.The ADR ranges progress from lower to upper bounds as running success exceeds the ADR threshold.

J Qualitative Analysis of Emergent Grasping Behaviors

Qualitative analysis shows that reposing pretraining produces natural multi-contact grasps, but zero-shot transfer can yield task-misaligned grasps that complete reposing yet fail insertion. ADEPT-trained downstream teachers retain more natural finger gaits than teachers trained from scratch.

  • Zero-shot reposing teacher: Reposing pretraining produces natural multi-contact closure, with fingers wrapping around the peg, but zero-shot grasps are not always task-aligned.These observations come from the reposing teacher applied zero-shot at ADR level 20.
  • Zero-shot reposing teacher: Bottom-up, upside-down grasps can solve reposing but fail peg insertion, which requires grasping the peg from above with its leg.The hand remains below the peg in this common failure mode, preventing completion of insertion.
  • Downstream-task teacher comparison: Downstream-task teachers trained from scratch lack the natural finger gaits that emerge from reposing pretraining.Figure 12 compares scratch-trained teachers at ADR level 50 with ADEPT-trained teachers at the same ADR level.
Loading 2608.19182v1…