Particle Filtering with Generative Trajectory Models for State Estimation and Control
Organizations: Laboratory for Information and Decision Systems, Massachusetts Institute of Technology, Cambridge, MA.
Abstract
Generative models can support both state estimation and control in robotics. For control, a common approach is to use the model as a policy that directly samples actions conditioned on observation histories. However, models that jointly represent states, actions, and observations can also predict how the robot state and its environment evolve. We investigate whether these predictions can improve state estimation and control through a change in inference procedure, without modifying the pretrained model. To this end, we use particle filtering to maintain competing hypotheses about state histories. A single pretrained diffusion model generates these hypotheses along with their predicted observations and associated action proposals. As new measurements arrive, particles are reweighted according to how well their predicted observations match the measurements. While particle filtering conventionally uses these weights only for state estimation, we also use them to guide action selection, favoring proposals associated with hypotheses that better explain the observations. Simulation experiments in partially observed navigation and vision-based manipulation demonstrate improvements in both state estimation and task performance compared to action generation without filtering, using the same pretrained model and no additional test-time information.
Figures & tables
| Method | Retained histories | Accum. weights | Final dist | ADE | RPE | |||
|---|---|---|---|---|---|---|---|---|
| Diffusion Policy | 1 | – | ||||||
| One-step selection | 1 | No | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-44\%,-38\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-51\%,-47\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-43\%,-40\%]} | |||
| Multi-history sel. | 32 | No | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-37\%,-32\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-46\%,-40\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-24\%,-15\%]} | |||
| Full PF | 32 | Yes | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-56\%,-49\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-62\%,-59\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptscriptstyle[-49\%,-45\%]} | |||
| Method | Final dist | ADE | RPE | |||
|---|---|---|---|---|---|---|
| Baselines: | ||||||
| Diffusion Policy | ||||||
| Ensemble | {\color[rgb]{0.5977,0.4023,0}\scriptstyle[-1\%,+10\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-16\%,-2\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-44\%,-40\%]} | |||
| PF w/ shuffled weights | {\color[rgb]{0.5977,0.4023,0}\scriptstyle[-6\%,+1\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-15\%,-8\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-40\%,-37\%]} | |||
| PF (weighted mean) | {\color[rgb]{0.5977,0.4023,0}\scriptstyle[-7\%,+4\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-22\%,-9\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-44\%,-41\%]} | |||
| With privileged information: | ||||||
| Training Dataset Size | ||||||
|---|---|---|---|---|---|---|
| Method | 1,000 | 10,000 | 100,000 | |||
| Diffusion Policy | ||||||
| Particle Filter | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-13\%,-6\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-13\%,-4\%]} | {\color[rgb]{0.707,0.1367,0.0938}\scriptstyle[+4\%,+11\%]} | |||
| PF w/ GT oracle | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-56\%,-49\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-52\%,-44\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-42\%,-33\%]} | |||
| Method | Final dist | ADE | RPE | |||
|---|---|---|---|---|---|---|
| Diffusion Policy | ||||||
| Particle Filter | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-11\%,-5\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-8\%,-4\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-15\%,-13\%]} | |||
| PF w/ pose oracle | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-48\%,-38\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-46\%,-39\%]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-31\%,-27\%]} | |||
| Method Details | Wrist-Camera-Only Transport | Wrist-Camera-Only ToolHang | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Method | Action selection | Num. particles | Persistent state? | SR (%) | State RMSE | SR (%) | State RMSE | ||||
| Diffusion Policy (DP) | – | 1 | |||||||||
| DP w/ persistent state | – | 1 | {\color[rgb]{0.5977,0.4023,0}\scriptstyle[-7.2,+0.2]} | {\color[rgb]{0.707,0.1367,0.0938}\scriptstyle[+7\%,+24\%]} | {\color[rgb]{0.5977,0.4023,0}\scriptstyle[-4.2,+5.4]} | {\color[rgb]{0.707,0.1367,0.0938}\scriptstyle[+7\%,+28\%]} | |||||
| DP ensemble | uniform mean | 32 | {\color[rgb]{0.5977,0.4023,0}\scriptstyle[-5.4,+5.0]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-29\%,-19\%]} | {\color[rgb]{0.707,0.1367,0.0938}\scriptstyle[-22.0,-13.2]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-37\%,-12\%]} | |||||
| DP ensemble w/ persistent state | uniform mean | 32 | {\color[rgb]{0.5977,0.4023,0}\scriptstyle[-4.8,+5.0]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-28\%,-13\%]} | {\color[rgb]{0.707,0.1367,0.0938}\scriptstyle[-19.2,-8.8]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-38\%,-2\%]} | |||||
| PF w/ weighted action averaging | weighted mean | 32 | {\color[rgb]{0.5977,0.4023,0}\scriptstyle[-0.2,+7.2]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-41\%,-28\%]} | {\color[rgb]{0.707,0.1367,0.0938}\scriptstyle[-16.0,-7.0]} | {\color[rgb]{0.0859,0.5,0.2344}\scriptstyle[-38\%,-15\%]} | |||||
Robotic Control
Robotic control studies how machines use sensing and computation to produce reliable movement and interaction in changing environments.
See all 626 papers