DexFLEX: Contact-Aware Foundation Controller for Command-Guided Dexterity

DexFLEX overview: a fingertip-motion draft from teleoperation or a learned policy is re-decided by DexFLEX into a contact-stable action, with real-world teleoperation and policy-learning results. — **DexFLEX** takes an upstream fingertip-motion *draft* (from teleoperation or a learned policy) and re-decides its contact behavior, emitting contact-stable joint commands. The same frozen controller boosts real-world teleoperation and serves as a low-level action interface for visuomotor policy learning.

78.3%

real-world teleoperation success
(up from 29.2% raw teleop)

2.0×

contact stability: normalized holding time (0.42 → 0.85)

46.7%

avg visuomotor policy-learning success

3.5×

over direct joint-action prediction

Abstract

Dexterous robot hands can receive useful motion intent from teleoperation or learned policies, but successful execution also depends on contact decisions that these commands rarely specify. We introduce DexFLEX, a contact-aware foundation controller that turns upstream fingertip-motion drafts into contact-consistent joint commands. Instead of treating a draft as a trajectory to copy, DexFLEX treats it as evidence about intent: from the current tactile-proprioceptive state, it proposes feasible short-horizon motion chunks, predicts their contact consequences, and selects the candidate that best follows the command while preserving future contact stability. At inference time, a simple Draft-Dream-Select procedure combines pure-prior proposals for recovery with draft-seeded proposals for responsiveness, then decodes the selected motion into executable joint-space targets. Across degraded-command simulation, real-world shared control, and visuomotor policy learning, the same trained controller improves robustness without retraining, increasing real-world teleoperation success from 29.2% with raw teleop to 78.3%, and reaching 46.7% policy-learning success, 3.5× direct joint-action prediction, and 1.56× prior-only correction.

Overview Video

A tour of DexFLEX: the contact-proprioceptive state, the Draft-Dream-Select loop, and results across simulation, real-world teleoperation, and policy learning.

This video contains audio.

Method

DexFLEX treats a draft not as a trajectory to copy but as evidence about intent: it proposes feasible short-horizon motions, predicts their contact consequences, and selects the candidate that best follows the command while keeping future contact stable.

The DexFLEX architecture: state encoding from proprioception and tactile history, a state-conditioned motion prior, a latent world model predicting contact consequences, consequence-aware scoring, and a contact-aware action decoder. — **The DexFLEX pipeline.** From a tactile-proprioceptive state, a consistency-flow motion prior proposes short-horizon fingertip-motion chunks (pure-prior and draft-seeded). A latent world model predicts each chunk's contact consequence; a contrastive scorer ranks them; the winner is decoded to joint-space targets by an inverse-dynamics decoder, looped at a receding horizon.

Click the diagram to enlarge.

01

State-conditioned motion proposals

We encode proprioception and per-fingertip FlexiTac tactile signals into a contact-proprioceptive state, then use a consistency-flow motion prior to propose feasible short-horizon fingertip-motion chunks, using a pure-prior branch for recovery and a draft-seeded branch for responsiveness.

02

Consequence-aware scoring

A latent world model rolls out the contact consequence of each candidate chunk, and a learned contrastive scorer ranks candidates by how faithfully they follow the command while preserving contact stability.

03

Draft-Dream-Select execution

At inference we form a candidate set from both branches, dream their consequences, select the best chunk, and decode it to executable joint-space targets through an inverse-dynamics decoder, repeating at a receding horizon.

Contact-Aware Motion Primitives

The motion prior captures a vocabulary of contact-rich behaviors (grasping, regrasping, releasing, reorienting, striking, and scooping) that DexFLEX recombines on the fly to honor a command without losing the object.

Robustness to Degraded Commands

We corrupt the command stream four ways (additive noise, command scaling, delay, and support-finger removal), then measure whether the controller still recovers the operator's intent instead of merely damping motion.

Robustness under command degradation: stable rollout duration, task progress, and intent recovery error across five corruption conditions and five methods. — **Robustness under command degradation.** Across five corruption conditions and five methods, DexFLEX best preserves rollout duration and task progress while minimizing intent-recovery error, most dramatically under **support-finger removal**, where naive methods drop the object. Baselines: Raw Command, Filter, DexGen, and DexFLEX w/o Tactile.

Expert under command noise ✗
motion degrades, contact is lost

DexFLEX, same corruption ✓
recovers intent, keeps contact stable

Shared Control on Real Hardware

Same operator, same task, same teleop interface. DexFLEX re-decides the contact behavior in the loop, turning a dropped hammer into a stable, repeatable strike.

×4

Raw teleop ✗ Hand slips, hammer drops

×4

DexFLEX ✓ Contact-stable strike

Six Tool-Use Tasks

Across six contact-rich tool-use tasks, DexFLEX raises average teleoperation success from 29.2% with raw teleop to 78.3%, and nearly doubles normalized holding time from 0.42 to 0.85.

Real-world teleoperation results: per-task success rate (SR, %) and averages.
Method	Brush	Nut	Hammer	Screwdrv.	Spatula	Toast	Avg SR ↑	Avg CT ↓	Avg NHT ↑
Raw Teleop	50	30	40	30	15	10	29.2	27.5	0.42
Filter	60	40	50	40	25	20	39.2	33.8	0.55
DexGen	70	55	60	55	40	35	52.5	31.2	0.65
DexFLEX w/o Tactile	75	60	65	60	50	45	59.2	30.8	0.71
DexFLEX	90	80	85	80	70	65	78.3	32.0	0.85

SR = success rate (%, higher is better) over 20 trials per task across two operators. CT = completion time (s, over successful trials). NHT = normalized holding time, the fraction of the horizon with a stable functional grasp (higher is better).

Selective Intervention

DexFLEX stays close to the operator most of the time and intervenes only when it matters: its correction magnitude spikes precisely at release, impact, scoop, and regrasp, and stays low otherwise.

Hardware. 16-DoF LEAP hand on a 7-DoF xArm; teleoperation via a Rokoko glove with a Vive tracker for wrist pose; one 10×6 FlexiTac tactile sensor per finger.

Policy Learning with a Frozen Interface

Without any retraining, DexFLEX serves as the action interface for a visuomotor imitation policy, absorbing contact decisions so the policy can focus on what to do, not how to hold.

Policy learning with different action interfaces: success on Pour Cup, Place Hammer, and Assemble Table for Direct Prediction, DexGen, and DexFLEX. — **Policy learning with different action interfaces.** Three contact-rich tasks, 50 demonstrations each, same image-based flow-matching policy; only the low-level action interface varies. With DexFLEX, success is **3.5×** direct joint-action prediction and **1.56×** a prior-only (DexGen) correction baseline. Baselines: Direct Joint Prediction, DexGen.

46.7%

average success across
the three tasks

3.5×

over direct
joint-action prediction

1.56×

over prior-only
correction (DexGen)

BibTeX

@inproceedings{dexflex2026anonymous,
  title     = {DexFLEX: Contact-Aware Foundation Controller
               for Command-Guided Dexterity},
  author    = {Anonymous Authors},
  booktitle = {Under review},
  year      = {2026},
  note      = {Paper under double-blind review}
}