Skip to content

Commit eb25f54

Browse files
DorsaRohsayakpaul
andcommitted
Add Diffusion Policy for Reinforcement Learning (#9824)
* enable cpu ability * model creation + comprehensive testing * training + tests * all tests working * remove unneeded files + clarify docs * update train tests * update readme.md * remove data from gitignore * undo cpu enabled option * Update README.md * update readme * code quality fixes * diffusion policy example * update readme * add pretrained model weights + doc * add comment * add documentation * add docstrings * update comments * update readme * fix code quality * Update examples/reinforcement_learning/README.md Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * Update examples/reinforcement_learning/diffusion_policy.py Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * suggestions + safe globals for weights_only=True * suggestions + safe weights loading * fix code quality * reformat file --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
1 parent 1a70054 commit eb25f54

File tree

2 files changed

+211
-1
lines changed

2 files changed

+211
-1
lines changed

examples/reinforcement_learning/README.md

+10-1
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,13 @@
1-
# Overview
1+
2+
## Diffusion-based Policy Learning for RL
3+
4+
`diffusion_policy` implements [Diffusion Policy](https://diffusion-policy.cs.columbia.edu/), a diffusion model that predicts robot action sequences in reinforcement learning tasks.
5+
6+
This example implements a robot control model for pushing a T-shaped block into a target area. The model takes in current state observations as input, and outputs a trajectory of subsequent steps to follow.
7+
8+
To execute the script, run `diffusion_policy.py`
9+
10+
## Diffuser Locomotion
211

312
These examples show how to run [Diffuser](https://arxiv.org/abs/2205.09991) in Diffusers.
413
There are two ways to use the script, `run_diffuser_locomotion.py`.
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,201 @@
1+
import numpy as np
2+
import numpy.core.multiarray as multiarray
3+
import torch
4+
import torch.nn as nn
5+
from huggingface_hub import hf_hub_download
6+
from torch.serialization import add_safe_globals
7+
8+
from diffusers import DDPMScheduler, UNet1DModel
9+
10+
11+
add_safe_globals(
12+
[
13+
multiarray._reconstruct,
14+
np.ndarray,
15+
np.dtype,
16+
np.dtype(np.float32).type,
17+
np.dtype(np.float64).type,
18+
np.dtype(np.int32).type,
19+
np.dtype(np.int64).type,
20+
type(np.dtype(np.float32)),
21+
type(np.dtype(np.float64)),
22+
type(np.dtype(np.int32)),
23+
type(np.dtype(np.int64)),
24+
]
25+
)
26+
27+
"""
28+
An example of using HuggingFace's diffusers library for diffusion policy,
29+
generating smooth movement trajectories.
30+
31+
This implements a robot control model for pushing a T-shaped block into a target area.
32+
The model takes in the robot arm position, block position, and block angle,
33+
then outputs a sequence of 16 (x,y) positions for the robot arm to follow.
34+
"""
35+
36+
37+
class ObservationEncoder(nn.Module):
38+
"""
39+
Converts raw robot observations (positions/angles) into a more compact representation
40+
41+
state_dim (int): Dimension of the input state vector (default: 5)
42+
[robot_x, robot_y, block_x, block_y, block_angle]
43+
44+
- Input shape: (batch_size, state_dim)
45+
- Output shape: (batch_size, 256)
46+
"""
47+
48+
def __init__(self, state_dim):
49+
super().__init__()
50+
self.net = nn.Sequential(nn.Linear(state_dim, 512), nn.ReLU(), nn.Linear(512, 256))
51+
52+
def forward(self, x):
53+
return self.net(x)
54+
55+
56+
class ObservationProjection(nn.Module):
57+
"""
58+
Takes the encoded observation and transforms it into 32 values that represent the current robot/block situation.
59+
These values are used as additional contextual information during the diffusion model's trajectory generation.
60+
61+
- Input: 256-dim vector (padded to 512)
62+
Shape: (batch_size, 256)
63+
- Output: 32 contextual information values for the diffusion model
64+
Shape: (batch_size, 32)
65+
"""
66+
67+
def __init__(self):
68+
super().__init__()
69+
self.weight = nn.Parameter(torch.randn(32, 512))
70+
self.bias = nn.Parameter(torch.zeros(32))
71+
72+
def forward(self, x): # pad 256-dim input to 512-dim with zeros
73+
if x.size(-1) == 256:
74+
x = torch.cat([x, torch.zeros(*x.shape[:-1], 256, device=x.device)], dim=-1)
75+
return nn.functional.linear(x, self.weight, self.bias)
76+
77+
78+
class DiffusionPolicy:
79+
"""
80+
Implements diffusion policy for generating robot arm trajectories.
81+
Uses diffusion to generate sequences of positions for a robot arm, conditioned on
82+
the current state of the robot and the block it needs to push.
83+
84+
The model expects observations in pixel coordinates (0-512 range) and block angle in radians.
85+
It generates trajectories as sequences of (x,y) coordinates also in the 0-512 range.
86+
"""
87+
88+
def __init__(self, state_dim=5, device="cpu"):
89+
self.device = device
90+
91+
# define valid ranges for inputs/outputs
92+
self.stats = {
93+
"obs": {"min": torch.zeros(5), "max": torch.tensor([512, 512, 512, 512, 2 * np.pi])},
94+
"action": {"min": torch.zeros(2), "max": torch.full((2,), 512)},
95+
}
96+
97+
self.obs_encoder = ObservationEncoder(state_dim).to(device)
98+
self.obs_projection = ObservationProjection().to(device)
99+
100+
# UNet model that performs the denoising process
101+
# takes in concatenated action (2 channels) and context (32 channels) = 34 channels
102+
# outputs predicted action (2 channels for x,y coordinates)
103+
self.model = UNet1DModel(
104+
sample_size=16, # length of trajectory sequence
105+
in_channels=34,
106+
out_channels=2,
107+
layers_per_block=2, # number of layers per each UNet block
108+
block_out_channels=(128,), # number of output neurons per layer in each block
109+
down_block_types=("DownBlock1D",), # reduce the resolution of data
110+
up_block_types=("UpBlock1D",), # increase the resolution of data
111+
).to(device)
112+
113+
# noise scheduler that controls the denoising process
114+
self.noise_scheduler = DDPMScheduler(
115+
num_train_timesteps=100, # number of denoising steps
116+
beta_schedule="squaredcos_cap_v2", # type of noise schedule
117+
)
118+
119+
# load pre-trained weights from HuggingFace
120+
checkpoint = torch.load(
121+
hf_hub_download("dorsar/diffusion_policy", "push_tblock.pt"), weights_only=True, map_location=device
122+
)
123+
self.model.load_state_dict(checkpoint["model_state_dict"])
124+
self.obs_encoder.load_state_dict(checkpoint["encoder_state_dict"])
125+
self.obs_projection.load_state_dict(checkpoint["projection_state_dict"])
126+
127+
# scales data to [-1, 1] range for neural network processing
128+
def normalize_data(self, data, stats):
129+
return ((data - stats["min"]) / (stats["max"] - stats["min"])) * 2 - 1
130+
131+
# converts normalized data back to original range
132+
def unnormalize_data(self, ndata, stats):
133+
return ((ndata + 1) / 2) * (stats["max"] - stats["min"]) + stats["min"]
134+
135+
@torch.no_grad()
136+
def predict(self, observation):
137+
"""
138+
Generates a trajectory of robot arm positions given the current state.
139+
140+
Args:
141+
observation (torch.Tensor): Current state [robot_x, robot_y, block_x, block_y, block_angle]
142+
Shape: (batch_size, 5)
143+
144+
Returns:
145+
torch.Tensor: Sequence of (x,y) positions for the robot arm to follow
146+
Shape: (batch_size, 16, 2) where:
147+
- 16 is the number of steps in the trajectory
148+
- 2 is the (x,y) coordinates in pixel space (0-512)
149+
150+
The function first encodes the observation, then uses it to condition a diffusion
151+
process that gradually denoises random trajectories into smooth, purposeful movements.
152+
"""
153+
observation = observation.to(self.device)
154+
normalized_obs = self.normalize_data(observation, self.stats["obs"])
155+
156+
# encode the observation into context values for the diffusion model
157+
cond = self.obs_projection(self.obs_encoder(normalized_obs))
158+
# keeps first & second dimension sizes unchanged, and multiplies last dimension by 16
159+
cond = cond.view(normalized_obs.shape[0], -1, 1).expand(-1, -1, 16)
160+
161+
# initialize action with noise - random noise that will be refined into a trajectory
162+
action = torch.randn((observation.shape[0], 2, 16), device=self.device)
163+
164+
# denoise
165+
# at each step `t`, the current noisy trajectory (`action`) & conditioning info (context) are
166+
# fed into the model to predict a denoised trajectory, then uses self.noise_scheduler.step to
167+
# apply this prediction & slightly reduce the noise in `action` more
168+
169+
self.noise_scheduler.set_timesteps(100)
170+
for t in self.noise_scheduler.timesteps:
171+
model_output = self.model(torch.cat([action, cond], dim=1), t)
172+
action = self.noise_scheduler.step(model_output.sample, t, action).prev_sample
173+
174+
action = action.transpose(1, 2) # reshape to [batch, 16, 2]
175+
action = self.unnormalize_data(action, self.stats["action"]) # scale back to coordinates
176+
return action
177+
178+
179+
if __name__ == "__main__":
180+
policy = DiffusionPolicy()
181+
182+
# sample of a single observation
183+
# robot arm starts in center, block is slightly left and up, rotated 90 degrees
184+
obs = torch.tensor(
185+
[
186+
[
187+
256.0, # robot arm x position (middle of screen)
188+
256.0, # robot arm y position (middle of screen)
189+
200.0, # block x position
190+
300.0, # block y position
191+
np.pi / 2, # block angle (90 degrees)
192+
]
193+
]
194+
)
195+
196+
action = policy.predict(obs)
197+
198+
print("Action shape:", action.shape) # should be [1, 16, 2] - one trajectory of 16 x,y positions
199+
print("\nPredicted trajectory:")
200+
for i, (x, y) in enumerate(action[0]):
201+
print(f"Step {i:2d}: x={x:6.1f}, y={y:6.1f}")

0 commit comments

Comments
 (0)