PaperlayerView exact revision on arXiv

Source-linked AI summary

CAPABLE: Capability-Aware Policy Adaptation via Behavioral Latent Encoding

Mohammad Khoshnazar, Mohammad Dehghani Tezerjani, Deyuan Qu, Zhiyuan Gao, Yanxiang Zhan, Jeroen Schafer, Andrew Melnik, Qing Yang, Michael Beetz

arXiv:2610.11971v1cs.ROcs.LGcs.SYeess.SY

TL;DR

Joint faults can change how commanded actions are physically executed, challenging frozen VLA policies and requiring recovery without relying on fault labels or explicit diagnosis. CAPABLE estimates actuator capability from command–response history and kinematics, then conditions bounded residual corrections. Across 28 LIBERO tasks, it improves success on a held-out actuator over a parameter-matched global-history baseline while preserving healthy performance.

  • Problem

    Frozen VLA policies assume the trained embodiment, while existing recovery approaches may require task-specific training, fault information, or privileged signals.

    Source: §I · p. 1, F. Baselines · p. 5

  • Method

    CAPABLE infers remaining actuator capability from shared per-joint command–response histories and live kinematics, then conditions bounded residual corrections to a frozen VLA.

    Source: §I · pp. 1–2, A. Per-Joint Command–Response History · p. 3

  • Results

    On 28 tasks, CAPABLE raises unseen-j2 success from 24.8% to 59.3%, a +17.4-point gap over Global-history SAC, while healthy success is 90.8% versus 91.4% for zero correction.

    Source: A. H1: Factorization Transfers to a… · p. 6

  • Takeaways & Limitations

    Across six leave-one-actuator-out splits, three of five unseen fault families, and hardware, CAPABLE outperforms a parameter-matched global-history baseline.

    Source: §IX · p. 7

    The study is simulation-heavy; hardware covers three task–joint conditions with software-enforced locks, and CAPABLE does not outperform Global-history SAC under partial effectiveness or range restriction.

    Source: §VIII · p. 7

Abstract

from arXiv · show

Vision-language-action (VLA) policies assume the embodiment on which they were trained and can fail when a joint fault changes how commanded actions are physically executed. Existing fault-recovery methods often require task-specific retraining, fault labels, explicit diagnosis, or privileged embodiment information. We introduce CAPABLE, a unified capability-aware adaptation framework for frozen VLAs that integrates self-supervised capability inference with residual reinforcement learning. CAPABLE infers capability, how much of the commanded motion each joint actually realizes and how that motion contributes to end-effector behavior, online from command-response history and kinematics using a temporal encoder shared across joints, Jacobian grounding, cross-joint attention, and self-supervised physical prediction. The resulting representation conditions a residual policy that adds bounded corrections to the VLA arm action without fault labels or faulty-joint identifiers. Across 28 LIBERO tasks, CAPABLE raises success on an actuator excluded from fault training from 24.8% to 59.3%, outperforming a parameter-matched global-history baseline by 17.4 points while preserving healthy performance. Leave-one-actuator-out experiments across six joints show that this transfer is not specific to one actuator, and additional evaluations characterize transfer to unseen fault families and demonstrate recovery on a physical Franka Panda. https://capable-vla.github.io/

I. INTRODUCTION

Joint faults can break the physical assumptions of frozen VLA policies, while stopping robots after faults can be costly. CAPABLE estimates each actuator’s remaining capability and uses it to adapt a frozen VLA, with evaluations spanning held-out actuators and a physical robot.

I. INTRODUCTION

  • Joint faults can make a VLA’s otherwise task-correct actions fail physically, creating reliability and safety concerns when stopping for repair is costly.The paper identifies incidents in space, surgical, and industrial robotics.Source: §I · p. 1
  • CAPABLE represents faults by each actuator’s remaining capability: how much commanded motion it realizes and how that motion affects the end effector.Capability is inferred from command–response history and kinematics without fault labels.Source: §I · pp. 1–2
  • CAPABLE keeps the VLA frozen and uses bounded residual reinforcement learning to correct its arm action without fault labels or faulty-joint identifiers.Its shared temporal encoder, live Jacobian grounding, cross-joint attention, and self-supervised physical prediction support actuator transfer.Source: §I · p. 2
  • The evaluation spans 28 tasks, six held-out joints, five unseen fault families, and a physical robot, comparing factorization with a parameter-matched unfactorized encoder.The study also examines where transfer to unseen fault families holds and where it fails.Source: §I · p. 2

II. RELATED WORK

Prior adaptation and fault-recovery approaches address parts of robot-policy adaptation, but commonly assume unchanged physical capabilities or rely on task-specific training, fault information, or reference data. The problem formulation treats faults as unobserved changes in dynamics that must be inferred from execution history.

A. Adapting Pretrained Robot Policies

  • Pretrained-policy adaptation includes VLA fine-tuning, robustness training, and recovery-oriented reinforcement learning, while residual methods freeze the base policy and learn corrections.These methods generally assume the robot’s physical capabilities remain unchanged; CAPABLE conditions corrections on online capability estimates.Source: A. Adapting Pretrained Robot Policies · p. 2

B. Fault-Tolerant Control and Learned Adaptation

  • Classical fault-tolerant control diagnoses faults or replans around locked joints, while learned approaches infer dynamics changes but often use privileged fault information.The cited approaches include predefined passive tolerance, active reconfiguration, and fault-informed locomotion training.Source: B. Fault-Tolerant Control and Learned… · p. 2

C. Learning-Based Fault Recovery

  • Prior learned recovery ranges from task-trained recurrent control to methods using embodiment vectors or fault supervision; CAPABLE substitutes task reward and physical prediction for those signals.J-PARC uses healthy reference rollouts and fault-presence and affected-joint supervision; DEFT conditions on an explicit embodiment vector.Source: C. Learning-Based Fault Recovery · p. 2

D. Weight Sharing Across Actuators

  • Unlike morphology-sharing methods that generalize across bodies, CAPABLE shares a module across actuators within one body and grounds tokens in the live Jacobian.The cited recurrent, attention, and FiLM components are established techniques.Source: D. Weight Sharing Across Actuators · p. 2

III. PROBLEM FORMULATION

  • A persistent fault changes transition dynamics, is unobserved by the policy, and must be inferred from execution history; an unseen actuator’s failure is excluded from training.The LIBERO reward is the task-success predicate, and the VLA parameters remain fixed during residual training.Source: §III · p. 3

IV. CAPABLE: CAPABILITY-AWARE POLICY ADAPTATION VIA BEHAVIORAL LATENT ENCODING

CAPABLE combines per-joint command–response histories with current kinematics to infer actuator capability and generate residual corrections. Its joint tokens use a common feature structure without explicit joint identifiers or fault-specific inputs.

IV. CAPABLE: CAPABILITY-AWARE POLICY ADAPTATION VIA BEHAVIORAL LATENT ENCODING

  • CAPABLE combines shared per-joint history encoding, current kinematic grounding, and self-supervised physical prediction to infer capability and generate residual corrections.The framework keeps the VLA frozen and uses the inferred representation for adaptation.Source: §IV · p. 3

A. Per-Joint Command–Response History

  • Each 28-dimensional joint token uses the same feature structure, with joint identity represented geometrically through its live Jacobian column rather than an explicit joint embedding.The encoder, actor, and critics receive no lock target, fault-family label, or severity value.Source: A. Per-Joint Command–Response History · p. 3

B. Shared Temporal Encoder

CAPABLE shares a temporal encoder across joints, combines history with kinematic context across the arm, and trains its capability representation through physical prediction. This design links command–response history to the motions the impaired arm can produce.

B. Shared Temporal Encoder

  • One temporal encoder with shared weights processes each joint’s recent history, allowing patterns learned on one joint to be recognized on an actuator unseen during training.Each joint stores 16 recent tokens.Source: B. Shared Temporal Encoder · p. 3

C. Kinematic Query and Cross-Joint Composition

  • Cross-joint attention composes joint information, while the command–response history and Jacobian together describe each joint’s motion and its end-effector effect.Combining estimates across all seven joints describes motions the impaired arm can still produce.Source: C. Kinematic Query and Cross-Joint… · p. 3

D. Self-Supervised Physical Prediction

  • Self-supervised physical prediction uses local joint-displacement and global end-effector-displacement decoders, with a kinematic loss aligning predicted and observed end-effector motion.The loss maps predicted joint displacements through the Jacobian before comparison.Source: D. Self-Supervised Physical Prediction · p. 3

E. FiLM-Conditioned SAC and Gradient Routing

The capability representation conditions the residual actor through FiLM, while encoder learning uses physical-prediction and critic objectives rather than actor gradients.

E. FiLM-Conditioned SAC and Gradient Routing

  • The listed input components come from robot sensing, controller outputs, internal counters, or known kinematics.The actor and critics receive a normalized residual observation.Source: E. FiLM-Conditioned SAC and Gradient… · pp. 3–4
  • FiLM uses the capability representation to modulate the actor and critic’s intermediate features.The bounded arm correction is added to the frozen VLA action, while the gripper bypasses RL.Source: E. FiLM-Conditioned SAC and Gradient… · p. 4
  • The actor’s deterministic output starts at zero, initially reproducing the base VLA arm action exactly.The residual observation includes the measured arm state and measured end-effector position and orientation.Source: E. FiLM-Conditioned SAC and Gradient… · pp. 3–4
  • The capability encoder is optimized with physical-prediction loss and a critic loss weighted by 0.05.Actor-loss gradients are blocked at zcap, keeping encoder learning dominated by physical prediction.Source: E. FiLM-Conditioned SAC and Gradient… · p. 4

A. Benchmark, Simulator, and Base Policy

The benchmark uses a shared OpenVLA-OFT checkpoint in LIBERO and models joint faults as physical locks; the evaluation pool retains tasks meeting healthy-success and usable-fault screening criteria.

A. Benchmark, Simulator, and Base Policy

  • The base checkpoint is shared across tasks, and OpenVLA-OFT executes action chunks of length eight.The residual policy changes only the six continuous arm dimensions; physical Franka Panda evaluation is also conducted under software-enforced joint locks.Source: A. Benchmark, Simulator, and Base Policy · p. 4
  • A joint lock is a MuJoCo equality constraint on the Panda joint, not an action mask, so the arm remains dynamically coupled to the locked link.The condition changes rebuild the relevant environment path, and lock drift is monitored.Source: A. Benchmark, Simulator, and Base Policy · p. 4

B. Held-Out-Blind 28-Task Pool

  • The 28-task pool retains tasks with at least 80% healthy success and at least one usable seen fault.Twelve of the 40 LIBERO tasks fail the screening rule, leaving 46 usable task–fault cells.Source: B. Held-Out-Blind 28-Task Pool · p. 4
  • Failures on j2 are excluded from training, while all 16 tasks with 0% base-VLA success under j2 remain evaluation cases.The j2 screening results are used only for reporting.Source: B. Held-Out-Blind 28-Task Pool · pp. 4–5

C. Task Curriculum and Replay Distribution

Training balances task sampling and targets a 10% healthy-execution share; evaluation uses held-out states and fixed checkpoints, with unseen-j2 tests spanning five fault families.

C. Task Curriculum and Replay Distribution

  • Tasks are sampled uniformly, while healthy execution receives 10% probability and usable seen faults share the remaining 90% by screened headroom.Replay uses the same distribution; uniform sampling over all 74 task–condition pairs would allocate about 37.8% of updates to healthy data.Source: C. Task Curriculum and Replay… · p. 5

D. Training and Deterministic Evaluation

  • Each method trains for 1.4M environment steps per seed and is evaluated on reserved initial states using three independently initialized seeds.Within each seed, methods use the same 300 held-out episodes per task–condition cell; six leave-one-actuator-out splits evaluate zero-shot transfer.Source: D. Training and Deterministic Evaluation · p. 5

E. Cross-Fault Generalization Protocol

  • The same checkpoints trained on persistent locks are evaluated on unseen j2 under five additional fault families without retraining.These include increased damping, increased Coulomb friction, partial actuator effectiveness, restricted joint range, and late-onset lock; severities are set from seen-joint pilots without j2 outcomes.Source: E. Cross-Fault Generalization Protocol · p. 5

F. Baselines

CAPABLE is compared with a parameter-matched unfactorized history encoder and other baselines; on the 28-task benchmark, it improves unseen-j2 success over zero correction and Global-history SAC.

F. Baselines

  • Global-history SAC matches CAPABLE’s main training and evaluation conditions within 1% trainable parameters, but encodes concatenated joint tokens with a global recurrent model.Classical redundancy resolution estimates joint availability online from command–response mismatch and redistributes the desired Cartesian motion through a weighted damped Jacobian pseudoinverse.Source: F. Baselines · p. 5
  • DEFT reaches 62.8% with oracle fault and task-state information, 3.5 points above CAPABLE.That comparison is not matched on information: CAPABLE uses non-privileged runtime observations, controller signals, and known kinematics.Source: F. Baselines · p. 5, A. H1: Factorization Transfers to a… · p. 6
  • Table I compares method assumptions, with learned methods sharing the task pool, fault conditions, budget, and held-out states.DEFT receives goal configurations derived from simulator state and the current actuator-availability vector.Source: F. Baselines · p. 5, Table I · p. 6

A. H1: Factorization Transfers to a Held-Out Actuator

  • Across 28 tasks, CAPABLE achieved 59.3% unseen-j2 success, up from 24.8%, compared with 41.9% for Global-history SAC and 32.9% for classical redundancy resolution.The CAPABLE–Global-history gap was +17.4 points under matched capacity, data, optimization, and evaluation. Table II is the primary result.Source: A. H1: Factorization Transfers to a… · p. 6

B. H2: Transfer Extends to Unseen Fault Families

Across the merged evaluations, CAPABLE outperforms Global-history SAC in pooled physical-robot trials and across six held-out actuators, while transfer to unseen fault families depends on fault signature. Ablations support a coupled temporal–kinematic capability representation.

B. H2: Transfer Extends to Unseen Fault Families

  • CAPABLE leads Global-history SAC on persistent lock, damping, friction, and late-onset lock on unseen j2, but not on partial effectiveness or range restriction.Successful cases share suppressed motion relative to command; the other two fault families are graded or state-dependent.Source: B. H2: Transfer Extends to Unseen… · p. 6

C. H3: Transfer Generalizes Across Held-Out Actuators

  • CAPABLE outperforms Global-history SAC on all six held-out actuators, so its transfer advantage is not specific to j2.The comparison uses the same eight LIBERO tasks; the j2 entry is from that subset, not the primary 28-task result.Source: C. H3: Transfer Generalizes Across… · p. 6

D. Mechanism Ablations

  • On unseen j2, removing self-supervised loss costs 15.6 points, unsharing the encoder 13.5, cross-joint attention 13.2, Jacobian query 11.5, and zcap conditioning 9.7.A memoryless residual scores 39.4% on unseen j2, below Global-history SAC; equal-strata replay raises healthy success to 92.3% while reducing fault recovery.Source: D. Mechanism Ablations · p. 6

E. Hardware Validation

  • CAPABLE reaches 85.0% versus Global-history SAC’s 45.0% on seen-joint conditions and 70.0% versus 40.0% on unseen-j2; overall rates across 30 trials per method are 80.0% versus 43.3%, significantly different (p = 0.007).Individual conditions are not significant at this sample size.Source: E. Hardware Validation · p. 6

JOINT IS EXCLUDED FROM FAULT TRAINING AND EVALUATED

CAPABLE combines shared command–response encoding with live kinematic grounding, outperforming a parameter-matched global-history baseline across held-out actuators while revealing limits on transfer to some fault types.

JOINT IS EXCLUDED FROM FAULT TRAINING AND EVALUATED

VII. DISCUSSION

  • 39.4% for kinematics without history and 45.8% for separate per-joint encoders trail full CAPABLE at 59.3% on unseen j2.The ablations support combining shared temporal encoding with live kinematic grounding rather than relying on either alternative alone.Source: Table VI · p. 7, §VII · p. 7
  • A shared temporal function gives stalled command–response patterns a common interpretation across joints, while the live Jacobian grounds corrections in configuration-specific geometry.The shared representation transfers across actuators without making the resulting correction independent of the robot’s current configuration.Source: §VII · p. 7
  • Real-robot success is reported under persistent locks, with 10 trials per condition and software-enforced joint-reference locks.Source: Table VII · p. 7
  • Transfer is limited for partial actuator effectiveness and range restriction, whose graded and state-dependent signatures are not captured by the current formulation.The paper identifies extending inference to graded and state-dependent impairments as a next step.Source: §VII · p. 7

VIII. LIMITATIONS

  • The broad study is simulation-heavy; hardware evaluation covers three task–joint conditions, and leave-one-actuator-out evaluation uses eight rather than all 28 tasks.Training uses persistent locks only, and CAPABLE does not outperform global-history SAC under partial effectiveness or range restriction.Source: §VIII · p. 7

IX. CONCLUSION

  • Across 28 LIBERO tasks, all six leave-one-actuator-out splits, three of five unseen fault families, and hardware, CAPABLE outperforms a parameter-matched global-history baseline.The conclusion frames factorized command–response representations as a practical route to fault recovery for frozen VLAs without fault labels or affected-joint supervision.Source: §IX · p. 7
Loading 2610.11971v1…