← Projects
2026-05

RAID: Retrieval-Augmented Inverse Dynamics

Proposed and implemented a retrieval-augmented inverse dynamics head for robotic manipulation that achieves 6.5x lower action prediction error than direct MLP baselines with only 25 demonstrations, by combining a frozen world-model encoder with cross-attention over a demonstration memory bank.

PythonPyTorchDeep LearningRoboticsReinforcement LearningTransformers Complete
GitHub →
RAID: Retrieval-Augmented Inverse Dynamics

Problem

World models for robotics — systems that learn to “dream” future visual states from current observations — have shown strong results in games and locomotion but have largely failed at robot manipulation. The reason is a specific gap: a world model predicts where the robot ends up, but not what motor command gets it there. As Prof. Ken Goldberg (UC Berkeley AUTOLab) put it: “it’s very hard to infer the associated robot commands.”

This is the action-inference problem. Given a current visual state and a predicted next state, what 7-DOF action (dx, dy, dz, roll, pitch, yaw, grip) produces that transition? A direct neural network struggles because: (1) many action sequences can produce the same next state, (2) contact dynamics are discontinuous — tiny changes in grip angle cause completely different outcomes, and (3) friction and surface texture are invisible in the visual observation.

Approach

We built RAID (Retrieval-Augmented Inverse Dynamics), a lightweight head that sits on top of a frozen world model and solves the action-inference step using retrieval rather than purely parametric learning.

Architecture (three components):

1. Frozen world model (GR-1). Rather than train a world model from scratch (GPU-weeks of compute), we repurposed GR-1 — a vision Transformer pretrained on large-scale video to predict future image features. We dropped its action head entirely and used only its visual encoder (384-dim class-token features) and one-step prediction head to produce the current state f_t and dreamed next state f̂_{t+1}.

2. Memory bank. Built offline from training demonstrations: every transition (f_t, f_{t+1}, a_t) is encoded with the frozen GR-1 encoder and stored. At inference, the bank is queried by cosine similarity.

3. RAID decoder. Takes the (f_t, f̂_{t+1}) pair and produces the 7-DOF action through three sub-components:

At training time, prior dropout (p=0.4) and Gaussian jitter (σ=0.05) on the retrieved prior prevent the optimizer from copying retrieval too aggressively, forcing the trunk and gate to remain useful when retrieval is imperfect.

GRPO fine-tuning. To bridge offline imitation to actual task completion, we applied Group Relative Policy Optimization (GRPO) — sampling groups of G=4 rollouts, computing within-group relative rewards as advantage signal, with no critic network required. RAID’s retrieval prior naturally structures the action space for RL: instead of exploring from scratch, GRPO refines which retrieved demonstrations to weight.

Experiment scope: 171 training runs across 53 experimental cells, testing DINOv2 and SigLIP encoders, direct MLP, mean-pool RAID, cross-attention RAID, Transformer IDM, and Diffusion Policy IDM heads on RoboMimic and LIBERO-Spatial benchmarks across N=25 to N=200 demonstrations.

Results

Headline result (LIBERO-Spatial, GR-1 encoder):

ConditionN=25N=50N=100N=200
Direct Visual MLP0.8520.6370.5700.552
RAID Cross-Attention0.1310.1540.1690.171

6.5x lower MSE at 25 demonstrations. The advantage is sharpest in the low-data regime — exactly where parametric baselines have no signal to lean on — and persists through 200 demonstrations, though the gap narrows as the direct model gets more data.

Key finding — the encoder is load-bearing. The 6x improvement is not from retrieval alone. Early experiments with DINOv2 and SigLIP encoders showed no comparable gain even with a 10-task pooled memory bank. GR-1’s pretraining objective (predict future image features) produces exactly the temporal-delta signal that inverse dynamics needs. Without that signal, a larger memory bank doesn’t help.

Temporal stride matters. With DINOv2 at native 20Hz (consecutive frames), RAID ties the direct MLP. With stride-5 sampling (~250ms windows), RAID pulls ahead — the longer time gap makes visual transitions discriminable for retrieval. GR-1’s dreamed features amplify this effect intrinsically.

GRPO probe: 195 GRPO updates from the N=200 behavior-cloned RAID checkpoint achieved best mean reward 1.226 and intermittent 25% success rate (1 of 4 rollouts completing the task). This confirms RAID is compatible with online RL fine-tuning, but stable manipulation success requires larger rollout groups and more compute.

Broader implication: If retrieval and world-model imagination can be fully decoupled, robots could adapt to new tasks by swapping or expanding memory banks rather than retraining large end-to-end policies — a path toward true one-shot task transfer.