Vision-language-action (VLA) policies perform manipulation tasks using camera images and language instructions, but object slip or visual occlusion can disrupt execution. Recovery is typically learned during policy training or delegated to a monitor that triggers a scripted repair after detecting a failure. We introduce ForkCast, a predictive recovery framework that compares counterfactual futures to determine whether and how to intervene in a frozen VLA policy. At each decision point, ForkCast evaluates three candidates: executing the policy's nominal action chunk (Commit), adjusting the camera viewpoint (Reobserve), and replanning the arm trajectory (Replan). A candidate-conditioned world model forecasts the latent outcome of each candidate at a common future horizon, and a shared task-progress head scores the predicted outcomes. By evaluating the policy's own continuation alongside the repair candidates, a single comparison based on task progress determines both whether and how to intervene, without a failure detector, an intervention threshold, or updates to the VLA backbone. On five simulated RoboTwin 2.0 tasks with injected occlusion and object slip, ForkCast increases pooled success from 13.3% for the frozen π0.5 policy to 30.0% and transfers to two additional VLA backbones without updating ForkCast. After adaptation on real-robot candidate rollouts, ForkCast also improves success over the base policy under both clean and visibility-limited conditions. Code will be released to benefit the research community.
Figures & tables
Fig. 1: Predictive recovery with ForkCast. At a common prediction horizon, ForkCast ranks Commit and available Reobserve and Replan candidates. (a) Candidate-view acquisition and policy re-query precede scoring. Reobserve ranks highest; ForkCast adopts its selected viewpoint and executes the corresponding cached action chunk. (b) At the next decision, Commit ranks highest and the frozen VLA policy resumes. Grids show predicted endpoint latents; bars show relative progress. Green outlines mark selected candidates. Views and scores are schematic.
Fig. 2: Overview of ForkCast. Commit retains the nominal plan, Reobserve acquires views and re-queries the frozen policy before scoring, and Replan generates an arm correction. A frozen visual encoder encodes history once for all candidates. Candidate arm and camera motion, aligned to a common horizon Hf , condition a shared V-JEPA 2-AC predictor [ 16 ] . It forecasts M milestones, spaced h control steps apart ( Hf=Mh ). A shared progress head scores endpoint latents, and ForkCast executes the highest-scoring candidate. Future grids represent undecoded latents; robot views and scores are schematic.
Method
clean
slip
occ
pooled-OOD
Commit
53.3
10.0
16.7
13.3
Fixed Reobserve
36.7
10.0
23.3
16.7
Fixed Replan
70.0
33.3
10.0
21.7
Random
53.3
16.7
16.7
16.7
Direct progress regressor
56.7
23.3
16.7
20.0
ForkCast w/o camera cond.
70.0
26.7
16.7
21.7
TABLE I: Recovery and candidate-selection comparison. Closed-loop success rates (%) across five RoboTwin 2.0 tasks using the frozen π0.5 policy. Clean has no injected disturbance; slip and occ denote object displacement and visual occlusion. Pooled-OOD averages the two disturbed conditions. ForkCast w/o camera cond. is retrained without the camera-offset channel. † The oracle uses privileged outcome information and is reported separately from deployable methods. The 95% Wilson confidence intervals (CIs) for pooled-OOD ( n=60 ) are: ForkCast [19.9,42.5] , Oracle [21.3,44.2] , Fixed REOBSERVE [9.3,28.0] , ForkCast w/o camera [13.1,33.6] , Fixed REPLAN [13.1,33.6] , Direct [11.8,31.8] , Random [9.3,28.0] , and COMMIT [6.9,24.2] .
Fig. 3: Qualitative recovery examples. Head-camera sequences compare the base policy (upper row) with ForkCast (lower row) in (a) simulated cup placement under occlusion, (b) simulated box placement after object slip, and (c) real-world can insertion. Frames progress from left to right. ForkCast uses Reobserve followed by Replan in (a,c), and Replan in (b), before resuming Commit . Blue, orange, and green markers denote Commit , Reobserve , and Replan , respectively.
Method
clean
slip
occ
pooled-OOD
XR-0 [ 3 ]
16.7
0.0
0.0
0.0
XR-0 + ForkCast
56.7
30.0
13.3
21.7
RDT [ 2 ]
10.0
3.3
6.7
5.0
RDT + ForkCast
33.3
33.3
13.3
23.3
π0.5 [ 1 ]
53.3
10.0
16.7
13.3
π0.5 + ForkCast
76.7
33.3
26.7
30.0
TABLE II: Cross-backbone transfer in simulation. Success rates (%) on five RoboTwin 2.0 tasks with three frozen policy backbones. ForkCast trained on π0.5 branches is applied to RDT and XR-0 without updates. Pooled-OOD averages slip and occlusion.
Variant
Sens.
Eff. cos
Rank ρs
All three
Full model
2.60×
+.025
+.268
yes
Replica (retrained)
4.08×
+.018
+.161
yes
(a) Conditioning channels
NoCond
0.79×
+.001
+.216
no
NoState
2.55×
+.015
+.157
no
NoCamera
1.19×
+.001
+.185
no
TABLE III: Component ablations and candidate-selection diagnostics. (a,b) Conditioning and loss ablations on 45 held-out mining states, each variant retrained with the channel or loss removed. Sens., Eff. cos, and Rank ρs denote sensitivity, latent-effect cosine similarity, and Spearman progress correlation. Replica is an independently retrained full model. Bold marks a failed criterion. (c) Candidate-selection diagnostics from 30 evaluation episodes per method. Ndiff=16 counts common candidate sets with unequal realized progress and is the denominator of the best-set hit rate.
Fig. 4: Real-world platform and evaluation settings. (a) Dual-arm platform with head-camera pan/tilt and a vertical lift carrying the camera and both arm bases. (b) Sponge transfer and (c) can insertion, each shown under unobstructed (left) and restricted-visibility (right) conditions. A box wall restricts visibility of the sponge in (b), while a lid restricts visibility of the insertion region in (c). Numbered markers and arrows indicate object motion; translucent overlays show different motion stages.
Fig. 5: Real-world success rates with and without ForkCast. Bars report success rates (%) over 30 trials per method in each setting.