# Vision-Language-Action Jump-Starting for Reinforcement Learning Robotic Agents

Source: [Vision-Language-Action Jump-Starting for Reinforcement Learning Robotic Agents](<https://www.sota2.com/research/paper/69e06cfe7d0172140ec7c9a2>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-12T05:00:05.418Z

Authors: Angelo Moroncelli, Roberto Zanetti, Marco Maccarini, Loris Roveda

Published: 2026-04-15T11:17:54.000Z

## Abstract

Reinforcement learning (RL) enables high-frequency, closed-loop control for robotic manipulation, but scaling to long-horizon tasks with sparse or imperfect rewards remains difficult due to inefficient exploration and poor credit assignment. Vision-Language-Action (VLA) models leverage large-scale multimodal pretraining to provide generalist, task-level reasoning, but current limitations hinder their direct use in fast and precise manipulation. In this paper, we propose Vision-Language-Action Jump-Starting (VLAJS), a method that bridges sparse VLA guidance with on-policy RL to improve exploration and learning efficiency. VLAJS treats VLAs as transient sources of high-level action suggestions that bias early exploration and improve credit assignment, while preserving the high-frequency, state-based control of RL. Our approach augments Proximal Policy Optimization (PPO) with a directional action-consistency regularization that softly aligns the RL agent's actions with VLA guidance during early training, without enforcing strict imitation, requiring demonstrations, or relying on continuous teacher queries. VLA guidance is applied sparsely and annealed over time, allowing the agent to adapt online and ultimately surpass the guiding policy. We evaluate VLAJS on six challenging manipulation tasks: lifting, pick-and-place, peg reorientation, peg insertion, poking, and pushing in simulation, and validate a subset on a real Franka Panda robot. VLAJS consistently outperforms PPO and distillation-style baselines in sample efficiency, reducing required environment interactions by over 50% in several tasks. Real-world experiments demonstrate zero-shot sim-to-real transfer and robust execution under clutter, object variation, and external perturbations.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [LiftPegUpright on ManiSkill LiftPegUpright v2](<https://www.sota2.com/research/sota/liftpegupright-on-maniskill-liftpegupright-v2.md>): LiftPegUpright; ManiSkill LiftPegUpright v2; Success Rate at t\*: 91.5.

- [LiftPegUpright on ManiSkill LiftPegUpright v3](<https://www.sota2.com/research/sota/liftpegupright-on-maniskill-liftpegupright-v3.md>): LiftPegUpright; ManiSkill LiftPegUpright v3; Success Rate at t\*: 63.3.

- [PickCube on ManiSkill PickCube v1](<https://www.sota2.com/research/sota/pickcube-on-maniskill-pickcube-v1.md>): PickCube; ManiSkill PickCube v1; Success Rate (t\*): 91.7.

- [PickCube on ManiSkill PickCube v2](<https://www.sota2.com/research/sota/pickcube-on-maniskill-pickcube-v2.md>): PickCube; ManiSkill PickCube v2; Success Rate (t\*): 95.1.

- [PickPlaceCube on ManiSkill PickPlaceCube v2](<https://www.sota2.com/research/sota/pickplacecube-on-maniskill-pickplacecube-v2.md>): PickPlaceCube; ManiSkill PickPlaceCube v2; Success Rate (t\*): 65.9.

- [ManiSkill PokeCube v2](<https://www.sota2.com/research/sota/maniskill-pokecube-v2.md>): PokeCube; ManiSkill PokeCube v2; Success Rate at t\*: 81.4.

- [PushCube on ManiSkill PushCube OOD v2](<https://www.sota2.com/research/sota/pushcube-on-maniskill-pushcube-ood-v2.md>): PushCube; ManiSkill PushCube OOD v2; Success Rate at t\*: 94.4.

- [Lift Cube on Franka Panda Robot](<https://www.sota2.com/research/sota/lift-cube-on-franka-panda-robot.md>): Lift Cube; Franka Panda Robot; Success Rate: 70.

- [Pick & Place on Franka Panda Robot](<https://www.sota2.com/research/sota/pick-and-place-on-franka-panda-robot.md>): Pick-&-Place; Franka Panda Robot; Success Rate: 80.

- [Peg Reorientation on Franka Panda Robot](<https://www.sota2.com/research/sota/peg-reorientation-on-franka-panda-robot.md>): Peg Reorientation; Franka Panda Robot; Success Rate: 20.

- [LiftPegUpright on ManiSkill LiftPegUpright v1](<https://www.sota2.com/research/sota/liftpegupright-on-maniskill-liftpegupright-v1.md>): LiftPegUpright; ManiSkill LiftPegUpright v1; Success Rate (t\*): 80.3.

- [PickPlaceCube on ManiSkill PickPlaceCube v1](<https://www.sota2.com/research/sota/pickplacecube-on-maniskill-pickplacecube-v1.md>): PickPlaceCube; ManiSkill PickPlaceCube v1; Success Rate at t\*: 39.1.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.13733>)

[PDF](<https://arxiv.org/pdf/2604.13733.pdf>)
