DexForesight: Self-Distilled Foresight for Dexterous Vision-Language-Action Models

Official checkpoint release for DexForesight (ICLR 2027 submission). DexForesight distills the future trajectory already recorded in dexterous demonstrations into two complementary supervision signals for a causal π0.5 policy: a representation-level target (action-conditioned V-JEPA 2 future-latent prediction) and an action-level correction (privileged future flow distillation into a residual adapter). On the 11-task DexJoCo benchmark it improves the official π0.5 baseline by +10.1 points on average (62.6% vs 52.5%).

This repository is gated (manual approval). Request access with your intended use.

Repository layout

Each checkpoint is one directory named by experiment ID:

DF-<method>-<setting>-<task>/
  params/            # orbax PyTree checkpoint (openpi π0.5 tree)
  assets/            # normalization stats
  _CHECKPOINT_METADATA

• method ∈ full / align-only / flow-only / ctx-phys / ctx-vis
• setting ∈ ro (rand-obj) / rf (rand-full)
• task ∈ hammer-nail, pinch-tongs, pick-bucket, click-mouse, fold-glasses, water-plant, assembly-b, unlock-ipad-b, hanoi-b, microwave-b, photograph-b

Mapping to paper cells:

Paper table	Method column	Experiment ID prefix
Table 1 rand-obj	DexForesight (Ours)	DF-full-ro-
Table 1 rand-full	DexForesight (Ours)	DF-full-rf-
Table 2	Alignment Only	DF-align-only-ro-
Table 2	Flow Distillation Only	DF-flow-only-ro-
Table 3	Physical Future	DF-ctx-phys-ro-
Table 3	Visual Future	DF-ctx-vis-ro-

Table 2 "Full DexForesight" = Table 1 rand-obj checkpoints. Table 3 "Physical + Visual" shares the Flow-Distillation-Only checkpoints (identical values in the paper). DF-official-mt-compat covers the two multi-task cells that use the official DexJoCo multi-task backbone packaged in the DexForesight module structure (adapter output zeroed; behavior-equivalent).

Usage

Load with openpi (see the paper's appendix for the software stack):

from openpi.policies import policy_config
policy = policy_config.create_trained_policy(config, checkpoint_dir="DF-full-ro-hammer-nail/")

Evaluation protocol

50 episodes × 3 seeds (0/1/2) per cell, official asynchronous OpenPI server–client protocol, DexJoCo rand-obj/rand-full settings. Reported as mean ± std success rate.

Citation

@inproceedings{dexforesight2027,
  title={DexForesight: Self-Distilled Foresight for Dexterous Vision-Language-Action Models},
  author={Anonymous},
  booktitle={ICLR},
  year={2027}
}