Diffusion Imitation from Observation (DIFO)

• DIFO is a family of diffusion-based methods that transforms observations into structured representations for imitation learning from state-only data to full perception-to-action mappings.
• It employs techniques such as transition-level discrimination, partial diffusion inversion, and score matching to enhance robustness and sample efficiency across diverse tasks.
• Empirical studies report significant improvements in performance and generalization, unifying various imitation strategies while addressing challenges like visual variations and embodiment mismatches.

Diffusion Imitation from Observation (DIFO) denotes a family of diffusion-based formulations for imitation from observation and learning from observation in which diffusion dynamics are used to transform observations, score or classify transitions, or generate actions and joint configurations from demonstrations. Recent papers use the term in materially different ways: as a transition-level discriminator for state-only imitation (Huang et al., 2024), as partial diffusion inversion for robust visual imitation learning (Hu et al., 2024), as end-to-end RGB-to-joint imitation of human poses (Spisak et al., 28 Jan 2025), and as one-shot human imitation through retargeted initialization of a pre-trained diffusion policy (Park et al., 25 Jun 2025). A related score-matching line diffuses states rather than actions or images and aligns expert and learner occupancies without expert actions (Wu et al., 2024). The resulting literature is unified less by a single canonical objective than by a shared use of diffusion trajectories as structured intermediate spaces for invariance, density modeling, or action refinement.

1. Terminological scope and conceptual range

The phrase “Diffusion Imitation from Observation” is not yet standardized. In some works, “observation” means state-only demonstrations in a Markov decision process, with the objective of matching expert transition occupancies without access to expert actions (Huang et al., 2024). In others, “observation” means raw RGB input, and diffusion is used either to suppress nuisance appearance variation before behavior cloning (Hu et al., 2024) or to map a human image directly to robot joint values (Spisak et al., 28 Jan 2025). In yet another formulation, a human video is converted into a retargeted robot trajectory that is then projected onto the action manifold of a pre-trained diffusion policy (Park et al., 25 Jun 2025). The score-matching framework in “Diffusing States and Matching Scores: A New Framework for Imitation Learning” does not use the name DIFO explicitly, but it is presented as a clean instantiation of diffusion imitation from observation through state diffusion and score matching (Wu et al., 2024). This suggests that DIFO is best understood as a research program rather than a single algorithm.

Formulation	Diffused quantity	Representative papers
Transition-level LfO reward learning	Next state or state transition pair	(Huang et al., 2024, Wang et al., 2023)
Non-adversarial occupancy alignment	States along OU or DDPM-style diffusion	(Wu et al., 2024)
Observation-space invariance	Images or partially inverted image latents	(Hu et al., 2024)
End-to-end perception-to-action	Robot joint targets or action chunks	(Spisak et al., 28 Jan 2025, Park et al., 25 Jun 2025)
Related viewpoint or bridge formulations	First-person images or observation endpoints in an SDE	(Spisak et al., 2024, Liu et al., 8 Dec 2025)

A common misconception is that DIFO always means action-free imitation. That is accurate for transition-only and score-matching formulations, but not for all usages. Stem-OB trains from datasets of image-action pairs, DIRIGENt trains on paired human-robot pose data, and BridgePolicy explicitly uses action-labeled demonstrations while redefining DIFO as observation-anchored sampling (Hu et al., 2024, Spisak et al., 28 Jan 2025, Liu et al., 8 Dec 2025).

2. Observation-space DIFO: invariance, shared representations, and perspective alignment

An observation-space interpretation of DIFO is developed most explicitly in “Stem-OB: Generalizable Visual Imitation Learning with Stem-Like Convergent Observation through Diffusion Inversion” (Hu et al., 2024). The setting is visual imitation learning from observation in which a policy $\pi_\theta(a \mid o)$ is trained from source-environment image-action pairs and must generalize to target environments with changes in lighting, textures, color or finish, background, and camera noise, while task-relevant geometry remains stable. Stem-OB uses pretrained image diffusion models to perform partial diffusion inversion on training observations. The inversion is “stem-like”: low-level visual details progressively collapse while high-level scene structure remains, so observations that share geometry and affordances map closer in the inverted space. The policy is trained on these transformed observations but is deployed on the original raw observations, with no test-time inversion and virtually zero overhead (Hu et al., 2024).

The method is technically simple. With total diffusion steps $T$ and partial inversion depth $t$, each observation is transformed as

$$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$

with $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$ i.i.d. across $t$. On the reported tasks, performance peaks around $10$–$15$ steps out of $T=50$, and the ratio $t/T$ matters more than the absolute number of steps (Hu et al., 2024). The paper’s attribute-loss analysis formalizes the idea that fine-grained appearance differences become indistinguishable earlier than coarse semantic differences, and empirical evidence is consistent with that account: latent distances are smaller for intra-category than cross-category image pairs, and a human study with $T$0 participants found that intra-category pairs begin to confuse participants at approximately $T$1 of $T$2, while cross-category confusion rises only around $T$3 (Hu et al., 2024).

The reported gains are strongest under photometric and texture variation. In real-world tasks using a Franka Emika Panda and two RealSense D435i RGB cameras, Stem-OB improves average success rate by $T$4 over the best baseline across Cup-to-Plate, Duck-to-Bowl, Open Drawer, and Turn-on Faucet. On Open Drawer, for example, the reported generalization success is $T$5 for Stem-OB versus $T$6 for Org and $T$7 for SRM; on Turn-on Faucet it is $T$8 for Stem-OB versus $T$9 for Org, $t$0 for SRM, $t$1 for Mix, and $t$2 for RO (Hu et al., 2024).

A related but distinct front-end appears in “Diffusing in Someone Else’s Shoes: Robotic Perspective Taking with Diffusion” (Spisak et al., 2024). There, a conditional diffusion model translates third-person RGB demonstrations into first-person images or robot joint configurations, reducing viewpoint mismatch rather than appearance shift. On paired third-person/first-person data from NICOL simulation, the diffusion translator reports test-set image metrics of MSE $t$3, L1 $t$4, and SSIM $t$5, outperforming pix2pix and CycleGAN on that benchmark (Spisak et al., 2024). A plausible implication is that observation-space DIFO can target several nuisance gaps—texture, lighting, and viewpoint—before policy learning proper.

3. State-only DIFO: diffusion discriminators, diffusion rewards, and score matching

The formulation most directly named “Diffusion Imitation from Observation” in the learning-from-observation literature is the transition-based method of “Diffusion Imitation from Observation” (Huang et al., 2024). In this setting, expert demonstrations contain only state transitions $t$6, and the learner must infer a policy through online interaction. DIFO models the next state with a conditional DDPM, using the current state $t$7 and a binary condition $t$8 indicating expert or agent. For a sampled diffusion step $t$9 and noisy version of $$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$0, the denoising loss is

$$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$1

Expert- and agent-conditioned losses are then converted into a classifier,

$$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$2

and the policy is trained with the GAIL-style transition reward

$$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$3

The discriminator is stabilized by an expert-only diffusion MSE anchor and can be used with SAC or PPO (Huang et al., 2024).

Empirically, this DIFO framework is reported to match or outperform the best baseline across navigation, locomotion, manipulation, and game domains, while exhibiting lower variance across seeds than GAIfO, AIRLfO, and WAIfO (Huang et al., 2024). On AntMaze, the data-efficiency study shows that with $$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$4 expert trajectories, all baselines fail at approximately $$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$5 success while DIFO retains approximately $$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$6 success. The method also remains robust under heavy action noise in a stochastic AntMaze variant (Huang et al., 2024).

DiffAIL is an adversarial predecessor with a different construction but a similar intuition (Wang et al., 2023). It models $$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$7 in state-action imitation and $$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$8 in state-only imitation with an unconditional DDPM, defines the discriminator directly as

$$\hat{o}^{t/T}_i = \sqrt{\bar{\alpha}_t}\, o_i + \sqrt{1-\bar{\alpha}_t}\,\tilde{\epsilon}_t,$$9

and uses a diffusion-derived surrogate reward averaged over timesteps. In the reported MuJoCo experiments, DiffAIL reaches or exceeds expert-level returns in several tasks, including HalfCheetah and Ant, and its discriminator generalizes better than GAIL to held-out expert trajectories (Wang et al., 2023).

A non-adversarial alternative is developed in “Diffusing States and Matching Scores: A New Framework for Imitation Learning” (Wu et al., 2024). Instead of training a classifier, the method diffuses states with an Ornstein–Uhlenbeck process,

$\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$0

trains score networks by denoising score matching, defines a Diffusion Score Divergence between learner and expert occupancy measures, and uses a score-derived cost for RL (Wu et al., 2024). Its main theoretical result gives first- and second-order instance-dependent bounds with linear scaling in the horizon, explicitly contrasting the method with compounding-error behavior typical of offline cloning. This distinguishes one important branch of DIFO from adversarial formulations: diffusion need not be a discriminator if it can instead provide a score-based discrepancy over occupancies (Wu et al., 2024).

4. Perception-to-action DIFO and cross-embodiment imitation

A second major branch uses diffusion to generate robot actions or joint configurations directly from observations of a demonstrator. In “DIRIGENt: End-To-End Robotic Imitation of Human Demonstrations Based on a Diffusion Model” (Spisak et al., 28 Jan 2025), DIFO means learning a direct perception-to-action mapping from a single RGB frame of a human demonstrator to a robot joint vector $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$1, without an explicit hand-engineered human–robot joint mapping. The architecture is a U-Net-style conditional diffusion model with attention-augmented convolutional blocks and a differentiable forward-kinematics head. The noisy target path perturbs robot joint targets at a random diffusion step, while the image path encodes the human image; the network uses $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$2-prediction and combines joint-space and Cartesian end-effector MSE losses (Spisak et al., 28 Jan 2025).

The key data-collection strategy reverses imitation: humans imitate slow, predictable robot arm motions, thereby creating natural human–robot pose pairs. The DIRI dataset contains approximately $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$3 paired samples; NICOL has $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$4 joint values, and the model is trained with Adam for $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$5 epochs, batch size $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$6, and $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$7 diffusion steps (Spisak et al., 28 Jan 2025). On the DIRI random $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$8 split, reported axis errors are $\tilde{\epsilon}_t \sim \mathcal{N}(0,I)$9 mm, $t$0 mm, and $t$1 mm on $t$2, with Cartesian MSE $t$3 and joint MSE $t$4; on the held-out-human $t$5-fold evaluation, average axis errors increase to $t$6 cm, $t$7 cm, and $t$8 cm, reflecting the difficulty of unique human-to-robot mappings. Runtime is approximately $t$9 ms per frame on an RTX 3060 (Spisak et al., 28 Jan 2025).

“DemoDiffusion: One-Shot Human Imitation using pre-trained Diffusion Policy” (Park et al., 25 Jun 2025) addresses a different cross-embodiment regime. Here the input is a single human video demonstration, from which 3D hand poses are reconstructed and kinematically retargeted into a rough open-loop robot trajectory $10$0. Rather than train a new paired model, DemoDiffusion injects this retargeted trajectory into a pre-trained generalist diffusion policy by initializing denoising at an intermediate step $10$1:

$10$2

Closed-loop denoising under current observations then projects the retargeted trajectory onto the manifold of plausible robot actions (Park et al., 25 Jun 2025).

The reported real-world results are large. Across eight tabletop tasks on a Franka Panda with a Robotiq gripper, average success is $10$3 for Pi-0 alone, $10$4 for kinematic retargeting, and $10$5 for DemoDiffusion. In dexterous simulation, success rates are $10$6 for the base robot policy, $10$7 for kinematic retargeting, and $10$8 for DemoDiffusion (Park et al., 25 Jun 2025). The paper explicitly interprets $10$9 as the control knob for the trade-off between adherence to the human demonstration and reliance on the learned robot prior.

A later related extension, BridgePolicy, redefines the sampling problem itself by embedding the observation inside the diffusion SDE and using an observation-informed endpoint rather than Gaussian noise (Liu et al., 8 Dec 2025). Although it still relies on action-labeled demonstrations during training, its DIFO interpretation is that inference should start “from what you see” rather than from an uninformed prior. This broadens the perception-to-action branch from conditional denoising to observation-anchored stochastic dynamics.

5. Shared mathematical motifs across formulations

Despite their diversity, DIFO methods repeatedly use the same diffusion primitives in different algorithmic roles. The first is standard DDPM forward noising,

$15$0

which appears directly in transition models, action generators, and observation inversion. Stem-OB repurposes this expression as a partial inversion operator on images rather than as a latent from which to sample a reconstruction; its key claim is that semantically similar images converge early in the inverted space while fine appearance details are swept away (Hu et al., 2024).

The second is conditional denoising toward a clean action or joint target. DIRIGENt uses $15$1-prediction, minimizing

$15$2

with an additional forward-kinematics loss on end-effector pose (Spisak et al., 28 Jan 2025). DemoDiffusion does not retrain the policy objective, but it modifies the reverse process by starting from a noisy version of a retargeted action chunk rather than pure noise (Park et al., 25 Jun 2025). BridgePolicy goes further by specifying endpoints $15$3 and $15$4 and learning reverse-time updates for an observation-embedded diffusion bridge (Liu et al., 8 Dec 2025).

The third is the use of diffusion losses or scores as discriminative signals. In the transition-level DIFO formulation, the expert–agent loss difference becomes a calibrated classifier over transitions (Huang et al., 2024). In the score-matching framework, forward-diffused states define a Diffusion Score Divergence,

$15$5

and policy learning is driven by a score-derived surrogate cost rather than a GAN-style discriminator (Wu et al., 2024). Collectively, these constructions show that diffusion in DIFO is not tied to a single modeling choice: it may be an invariant representation builder, a conditional generator, a classifier surrogate, or a score-based divergence estimator.

6. Empirical profile, limitations, and open problems

Across the literature, DIFO methods tend to report their strongest gains in settings with multimodality, nuisance visual variation, or embodiment mismatch. Stem-OB is particularly effective in photorealistic simulation and real-world robotics, but it does not dominate in most MimicGen settings, where limited textures and low resolution reduce the benefit of inversion (Hu et al., 2024). Transition-based DIFO improves stability and sample efficiency in state-only continuous-control LfO, especially under stochasticity and limited demonstrations (Huang et al., 2024). DIRIGENt generalizes across tasks and robots but incurs larger errors on held-out demonstrators because human–robot mappings are demonstrator-dependent (Spisak et al., 28 Jan 2025). DemoDiffusion performs well in natural scenes with distractors and object-placement variation because denoising remains conditioned on current observations, but it still depends on accurate hand tracking, camera calibration, and a sufficiently capable base policy (Park et al., 25 Jun 2025).

Several limitations recur. Stem-OB explicitly notes that large structural scene changes, extreme occlusions, motion blur, and temporal instability remain problematic, and that less photorealistic domains with minimal textures may not benefit (Hu et al., 2024). DIRIGENt highlights depth ambiguity from single RGB frames, persistent inverse-kinematics redundancy, and the absence of explicit collision, self-collision, or torque limits in training (Spisak et al., 28 Jan 2025). The transition-level DIFO paper notes that rewards based only on $15$6 ignore action-dependent dynamics and assume a shared state space between expert and learner (Huang et al., 2024). The score-matching framework emphasizes compute and sensitivity to the diffusion schedule, especially if one wishes to scale from low-dimensional states to raw visual observations (Wu et al., 2024).

Open directions in the papers are correspondingly diverse. Stem-OB proposes conditional diffusion and ControlNet, temporal consistency across video frames, online adaptation, and evaluation with more visual IL backbones (Hu et al., 2024). DIRIGENt proposes trajectory diffusion with temporal conditioning, stronger 3D conditioning, classifier-free guidance, multi-robot transfer, and explicit safety regularizers (Spisak et al., 28 Jan 2025). DemoDiffusion points to better IK solvers, contact-aware retargeting, explicit guidance by the retargeted trajectory, multiple demonstrations, and adaptive selection of the initialization step $15$7 (Park et al., 25 Jun 2025). BridgePolicy suggests that observation-embedded bridges may be a general recipe for strengthening perception–action coupling when conditioning alone is insufficient (Liu et al., 8 Dec 2025).

The broader significance of DIFO lies in this plurality. Some formulations remove the need for expert actions, some remove the need for paired human–robot data, some remove the need for test-time preprocessing, and some instead improve robustness or embodiment transfer without eliminating supervision entirely. The central question is therefore not whether DIFO denotes one algorithm, but which part of the imitation pipeline diffusion is restructuring: the observation space, the occupancy discrepancy, the action prior, or the stochastic dynamics linking perception to control.