Learning from Relative Action Credit for Escaping Local Minima

Post-training of vision-language-action (VLA) models typically relies on expert demonstrations and policy interaction trajectories. However, in long-horizon manipulation, a single episode often spans hundreds of control steps and multiple phases, while success or failure is only revealed at episode termination. Policy improvement therefore requires step-level credit signals to distinguish behaviors that advance the task from those that stall or regress. We present PACE, a credit-assignment framework for post-training on long-horizon manipulation, centered on a phase-progress-aware critic. PACE consists of two key modules: (1) the Global-Local Cooperative Value-Correction Critic (GLC-Critic) aggregates visual and motion-difference features within local temporal windows to infer the phase and intra-phase progress of each step, and applies residual correction to a discretized remaining-cost distribution accordingly, enabling step-level credit assignment; (2) Progressive Policy Distillation (PPD) converts credit into positive and negative conditions via task-wise thresholds and trains a credit-conditioned action generation policy: it first protects the pretrained policy with high-credit positive samples, then incorporates all positive and negative credits to learn the quality boundary, and at inference amplifies high-credit behaviors through the difference between conditional outputs. Extensive simulation experiments and diverse real-world robotic-arm experiments demonstrate that PACE consistently achieves significant improvements over the strongest baseline.

Training agents directly on the task they will be deployed on often leads to the best policy performance, but training on physical robots can cause damage through falls, so minimizing training falls is critical. A standard mitigation hands control to a separate recovery policy whenever the agent leaves a designer-specified safe region. The resulting mixed-policy rollouts bias the on-policy update, and importance sampling cannot remove this bias when the recovery policy is deterministic. We propose SafeExplorer, a drop-in modification of proximal policy optimization (PPO) built on an unbiased policy-gradient estimator of the mixed-policy return. The estimator uses the score function only inside the safe region and never evaluates the recovery policy's density, so deterministic recovery policies pose no problem. On final reward, the estimator matches or outperforms truncated importance sampling, which is itself biased for that same gradient. SafeExplorer also speeds up learning near the safe-region boundary with a closed-form critic target at recovery-triggering states when dynamics and recovery are deterministic, and an imitation loss that copies recovery actions only from successful recoveries. On three simulated locomotion environments with ten seeds each, and given an already available recovery controller, SafeExplorer incurs $255\times$, $43\times$, and $21\times$ fewer training falls than standard PPO on HalfCheetah, Ant, and Unitree Go1. SafeExplorer also matches or exceeds the final reward of PPO. On Ant, where 8.7% of recovery interventions end in a fall, over $40\times$ the HalfCheetah rate, SafeExplorer reaches the success threshold on every seed and none of the five baselines reaches the threshold on any seed.

Action chunking provides temporal abstraction in reinforcement learning by selecting short action sequences instead of individual actions, but many existing approaches face two limitations in high-dimensional robotic control. First, many rely on value functions over action chunks, which can be difficult to learn as action dimensionality and chunk length grow. Second, executing chunks open-loop removes within-chunk feedback, limiting reactivity in contact-rich tasks. We present Action Chunking PPO (ACPPO), a PPO extension that uses a chunked actor while retaining a standard state-value critic, thereby avoiding chunked Q-functions. We further propose ACPPO-Corr, which augments the chunk planner with a stepwise feedback corrector that adjusts planned actions online within each chunk. Across 25 simulated robotics tasks from IsaacGym and Bi-DexHands, spanning locomotion, arm manipulation, and dexterous hand-object interaction, ACPPO-Corr achieves the strongest aggregate performance among evaluated methods and performs best on both decision-frequency-sensitive and decision-frequency-neutral task subsets. Ablations show that moderate chunk lengths work best and that corrector regularization is important for balancing chunk-level planning with local feedback. These results suggest that action chunking can be effective in online PPO when chunk-level planning is paired with closed-loop correction. The code is available at: https://github.com/hshhahn/ACPPO.