AI

CoRe Recovers a Frozen VLA by Imagining the Next Move

Robb Harlan 5 min read

When a VLA gets shoved mid-task, most stacks either fine-tune on failure tapes or hand the mess to a second model. CoRe (Counterfactual Realignment), posted August 14, keeps the policy frozen and plans the recovery in its head.

It watches the action-expert suffix latent, flags an out-of-distribution stretch, then rolls the same policy forward from a recent clean state on synthesized RGB. No extra world model. Physical motion starts only after that imagined continuation looks safe.

Real xArm6 pick, push, and place tasks used to test CoRe recovery
Physical xArm6 tasks. Source: Zhang et al., arXiv:2608.14822.

Detect, imagine, move less

Detection is a Mahalanobis score on the action head, calibrated on successful rollouts. Recovery anchors at the latest in-distribution frame that still keeps finished subgoals. From there CoRe synthesizes exo and ego views (voxel stores + forward-kinematics arm + inpainting), queries the frozen policy, and stops before contact.

A probe then asks which displaced objects actually matter. It restores the smallest subset whose imagined arm path stays close to a full-restore reference. The rest of the clutter stays put.

CoRe pipeline from OOD trigger through virtual planning to physical handoff
Virtual planning, then a short physical rearrange. Source: Zhang et al., arXiv:2608.14822.

The numbers they report

On LangSwitch (xArm6 in ManiSkill; mid-episode instruction changes), CoRe lifts three backbones from 18.3–35.5% to 82.6–92.7% average success. No false triggers in 1,500 unswitched runs.

On LIBERO-Long with injected arm collisions, grasp drops, and object shoves, CoRe reaches 69.8–88.2% versus 12.8–43.5% for the bare policies.

On a real xArm6 with two RealSense D435i cameras and a π0.5 backbone (100 demos per skill), 20 paired trials per condition:

ConditionBare+CoRe
Re-grounding (pick→pick)6/2018/20
Re-skilling (pick→push)1/2018/20
Conflict (place→push)1/2018/20
Arm collision9/2017/20
Scene rearrangement4/2019/20

Those re-skill and conflict jumps are +85 points, matching the abstract. Median virtual planning is 9.7 s. Median physical pick-and-place to test a hypothesis would have been 20.5 s of arm time. Scene-rearrangement recoveries restore 1.15 of 2.77 above-tolerance objects and leave 58% untouched. The abstract’s 42.2% drop in physical restorations is the same idea measured across the paper’s suite.

LangSwitch task families: re-grounding, re-skilling, and conflict
Language-switch families in simulation. Source: Zhang et al., arXiv:2608.14822.

Synthesized rollouts stay within the pre-contact margins they use for decisions (median TCP divergence 0.36–3.28 cm vs ground-truth renders). The method still assumes RGB-D, object tracks, and a planner that can pick the objects it decides to move.

A Human’s Take

I have watched too many “recovery” papers collect a new failure dataset and call it autonomy. Planning the retry on fake cameras, then moving only the cup that actually blocks the next motion, is closer to how a tired human cleans up a knock.

It is still a tabletop with tracked objects. The interesting bet is whether that suffix latent stays a decent smoke alarm once the scene is a kitchen, not four cubes and two plates.

Sources