RAFT: Reasoning-Aligned Fine-Tuning Improves Robustness of VLA Post-Training

Shresth Grover, Andrey Kolobov, Dhruv Shah

November 2026

RAFT rollouts on the real DROID and YAM robots under changed object colors, distractors, backgrounds and box poses.

TL;DR Before fine-tuning a VLA on a few demonstrations, RAFT trains LoRA adapters to predict VLM-generated reasoning annotations (boxes, keypoints, gripper traces), then freezes them. On a real DROID arm it succeeds in 75% of out-of-distribution trials versus 28% for plain fine-tuning, and on a bimanual YAM task it reaches 70% versus 36% for the best baseline.

What they do

Fine-tuning a generalist VLA on a few demonstrations of a new task raises success in the training setup but makes the policy brittle to new objects, backgrounds and phrasings. RAFT adds a step before action fine-tuning. The policy first learns task-relevant reasoning (which objects matter, where to grasp, how the gripper should move), and that knowledge stays frozen while it learns the actions.

How it works

Stage I, reasoning alignment. Gemini 3.1 Pro annotates demonstration frames with bounding boxes, keypoints, gripper traces, spatial descriptions or question-answer text. LoRA adapters in the vision encoder and language model learn to predict these annotations as tokens, with all other weights frozen.
Stage II, action adaptation. The adapters and the SigLIP vision encoder are frozen. The language model, action expert and projection layers are fine-tuned on the demonstrations, so action gradients never reach the adapters.
Policies. MolmoBot in simulation and on DROID, π0.5 on YAM.

Results

DROID pick and place (50 demos): 75.0% success over out-of-distribution trials versus 27.9% for plain fine-tuning (new object color 83 vs 33%, distractors 75 vs 25%, new tabletop 65 vs 25%). In the training setup the two are close (63 vs 60%).
Bimanual YAM open-box-and-place (100 demos): RAFT succeeds in 70.0% of trials pooled over all conditions, versus 36% for DeLock, 35% for plain fine-tuning and 29% for RETAIN. Plain fine-tuning usually opens the box but fails to place the object.
Simulated pick and place: 83.2% mean success versus 79.6% for plain fine-tuning. The order of training matters: co-training reasoning and actions reaches 63.2%, and leaving the adapter trainable in Stage II reaches 74.0%.

Limitations

The best annotation type depends on the task, and how to choose it is left open. In simulation the gains are modest, and with 25 demonstrations RAFT and plain fine-tuning tie at 46.2%. Open-drawer success stays low (44% in the training setup). Real-robot conditions use only 20-30 trials each.

Figures

From Figure 1: Plain fine-tuning on actions succeeds in the training layout but fails under small scene changes. RAFT... — From Figure 1: Plain fine-tuning on actions succeeds in the training layout but fails under small scene changes. RAFT first learns task-relevant reasoning in LoRA adapters, freezes them while fine-tuning on the same demonstrations, and succeeds under these shifts.

From Figure 2: Overview of RAFT. Stage I updates only the reasoning adapters, using reasoning annotations. Stage II f... — From Figure 2: Overview of RAFT. Stage I updates only the reasoning adapters, using reasoning annotations. Stage II freezes the adapters and updates the language model, action expert and projection layers on action demonstrations.

From Figure 6: Real-robot success for DROID pick and place, in the training setup and with a new object color, distra... — From Figure 6: Real-robot success for DROID pick and place, in the training setup and with a new object color, distractors or a new background.

From Figure 6: Real-robot success for the bimanual YAM task (open a box, place an object inside), in the training set... — From Figure 6: Real-robot success for the bimanual YAM task (open a box, place an object inside), in the training setup and under clutter, box pose, object color and background shifts.

Abstract

Adapting a generalist vision-language-action (VLA) policy to a target task with limited demonstrations can improve nominal task success while reducing robustness to variations in appearance, objects, and instructions. We introduce reasoning-aligned fine-tuning (RAFT), a two-stage post-training procedure: RAFT first trains low-rank adapters on task-relevant reasoning annotations, then freezes them while fine-tuning the policy on action demonstrations. On a physical DROID robot, RAFT achieves $75.0\%$ out-of-distribution success, compared with $27.9\%$ for plain fine-tuning; on a multi-stage bimanual YAM task, it reaches $70.0\%$ mean success, compared with $36.0\%$ for the strongest baseline. RAFT also improves success under visual, object, and instruction shifts in simulated pick-and-place and open-drawer tasks. Compared with plain fine-tuning, RAFT policies place more attention on annotated task regions and yield higher linear-probe accuracy for object identity in later layers. A broader reasoning-data mixture further improves downstream success across demonstration budgets.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong