PALM: Progress-Aware Policy Learning via Affordance Reasoning for Long-Horizon Robotic Manipulation

TL;DR

PALM tackles the challenge of long-horizon, language-conditioned robotic manipulation by introducing two core innovations: a fine-grained affordance predictor that forecasts four structured future cues (Global, Local, Spatial, Dynamic) and a progress-aware inverse-dynamics policy that jointly outputs actions and a continuous subtask progress value. It leverages large-scale pre-training on diverse robot and long-horizon video data, then fine-tunes on human-annotated affordance datasets to ground planning in real-world scenarios. Empirically, PALM achieves state-of-the-art results on CALVIN ABC→D (12.5% improvement) and LIBERO-LONG (91.8% success), with robust generalization under real-world disturbances and limited fine-tuning. The approach provides interpretable internal signals that connect perception, reasoning, and control, enabling stable, end-to-end long-horizon manipulation and offering practical benefits for safety, debugging, and deployment in open-world settings.

Abstract

Recent advancements in vision-language-action (VLA) models have shown promise in robotic manipulation, yet they continue to struggle with long-horizon, multi-step tasks. Existing methods lack internal reasoning mechanisms that can identify task-relevant interaction cues or track progress within a subtask, leading to critical execution errors such as repeated actions, missed steps, and premature termination. To address these challenges, we introduce PALM, a VLA framework that structures policy learning around interaction-centric affordance reasoning and subtask progress cues. PALM distills complementary affordance representations that capture object relevance, contact geometry, spatial placements, and motion dynamics, and serve as task-relevant anchors for visuomotor control. To further stabilize long-horizon execution, PALM predicts continuous within-subtask progress, enabling seamless subtask transitions. Across extensive simulation and real-world experiments, PALM consistently outperforms baselines, achieving a 91.8% success rate on LIBERO-LONG, a 12.5% improvement in average length on CALVIN ABC->D, and a 2x improvement over real-world baselines across three long-horizon generalization settings.

Figures (8)

• Figure 1: In contrast to vanilla VLAs that directly map inputs to actions or to predictive methods that forecast dense future images, PALM introduces learnable queries to forecast a structured set of future affordances. Conditioned on these affordances, a diffusion-based policy jointly decodes the robot's action and a continuous progress value, enabling temporal state tracking and seamless subtask transitions in multi-step, long-horizon tasks. Our training strategy utilizes large-scale pre-training on both robot datasets (BridgeDataV2 khazatsky2025droidlargescaleinthewildrobot, DROID walke2024bridgedatav2datasetrobot) and long-horizon video data (EPIC-KITCHENS damen2020epickitchensdatasetcollectionchallenges, RoboCerebra han2025robocerebralargescalebenchmarklonghorizon), followed by fine-tuning on our collected human-annotated affordance dataset. Consequently, PALM attains state-of-the-art performance on long-horizon simulation benchmarks (CALVIN ABC$\rightarrow$D mees2022calvinbenchmarklanguageconditionedpolicy, LIBERO-LONG liu2023liberobenchmarkingknowledgetransfer) and demonstrates strong results under real-world long-horizon generalization settings.
• Figure 2: PALM Overview.(a) Model Architecture: Given a language instruction $l$, observation $o_t$, and robot state $s_t$, PALM encodes each modality using frozen encoders to obtain text, visual, and state tokens. These tokens are fused by a GPT-style transformer with unidirectional attention and two specialized query sets: fine-grained affordance and action–progress. During training, affordance queries attend to context tokens to predict foresight $\hat{\mathbf{F}}_{t+n}$ with four supervised heads (<Global>, <Local>, <Spatial>, <Dynamic>) that ground future scene understanding. At inference, the affordance heads are removed; the action–progress query attends to both context and affordance foresight to condition a diffusion transformer that predicts action $\hat{a}_{t:t+n-1}$ and progress $\hat{p}_{t:t+n-1}$ trajectories for continuous control. (b) Structured Attention: Affordance subqueries attend only to shared context tokens to stay disentangled, while both query types use causal attention to preserve temporal consistency.
• Figure 3: Ablation studies of affordance components. on CALVIN ABC$\rightarrow$D and LIBERO-LONG benchmarks demonstrate the effectiveness of the four components of affordance prediction.
• Figure 4: Real-world experimental setup and task design.Left: We use a UFACTORY xArm6 robot with the matched Gripper G2 and two RealSense D455 cameras. Right: We design a real-world long-horizon manipulation task consisting of six consecutive subtasks, driven by a single high-level instruction.
• Figure 5: Random Relocation Disturbances. Predicted progress in the pick up grape subtask under two random grape relocations.