Source note
ReCAPA: Hierarchical Predictive Correction to Mitigate Cascading Failures
Summary
ReCAPA is a vision-language-action framework for long-horizon embodied tasks that tries to stop early mistakes from spreading through the rest of a plan. It adds correction at the action, subgoal, and full-trajectory levels, then reports gains on VisualAgentBench, MineDojo, and AI2-THOR.
Problem
- Long-horizon VLA agents often fail after one bad intermediate step, because local errors change later subgoals and actions and create cascading failures.
- Existing methods often rely on fixed task decompositions or post-hoc fixes, so they react late and may keep local steps aligned while drifting away from the overall task intent.
- This matters because embodied tasks such as navigation, manipulation, and crafting require many dependent decisions; the paper cites prior results where one subgoal error can cut later-step performance by over 60% on benchmarks such as VirtualHome and AI2-THOR.
Approach
- ReCAPA builds a three-level hierarchy over behavior: actions, subgoals, and full trajectories. Lower levels predict the representation of the next higher level, so the model can detect mismatch early.
- The main module, Hierarchical Predictive Correction (HPCC), uses Transformer predictors and an InfoNCE loss to make action segments predict subgoal embeddings and subgoal segments predict trajectory embeddings.
- It adds prompt-trajectory alignment in two ways: a Sinkhorn optimal-transport loss for global alignment between the whole trajectory and the instruction, and a score-field module that gives local corrective gradients for step-wise deviations.
- During inference, an LLM (GPT-4o-mini) proposes subgoals and completion criteria. ReCAPA then filters or reranks candidate actions using subgoal similarity and trajectory-level Sinkhorn consistency, with fallback rules if thresholds fail.
- The paper also introduces two diagnostics for failure dynamics: Error Propagation Rate (EPR), which measures how much one error raises later error probability, and Propagation Attenuation Coefficient (PAC), which measures how fast error effects decay.
Results
- On AI2-THOR, ReCAPA reaches SR 0.75, beating LLaMAR 0.68 and GPT-4V 0.66; it also reports TR 0.93 and Balance 0.93.
- On VisualAgentBench, ReCAPA reports 58.65 AVG, above GPT-4o mini 54.15, Gemini 2.5 Flash 53.00, and Claude-4-Sonnet 50.25. Per domain, it scores 50.6 on OmniGibson and 66.7 on Minecraft.
- The paper states relative gains of +5.65% on VisualAgentBench, +9% on MineDojo, and +7% on AI2-THOR over strong baselines.
- For error propagation, on OmniGibson at k=10, ReCAPA reports EPR_10 = 0.082, compared with about 0.3 for GPT-4o-mini and Gemini-2.5, and above 0.45 for Claude-4-sonnet.
- In ablations, removing HPCC drops Behavior SR from 72.2 to 59.3. On AI2-THOR, HPCC-AT reaches 0.73 SR and HPCC-ST 0.69, both above the action+subgoal-only variant. The paper says using Sinkhorn and Score-field together gives the best overall performance.
- The excerpt gives strong qualitative claims for MineDojo such as leading in 8 of 10 long-horizon tasks, but it does not include the full task table in the provided text.